Skip to content

GitLab

Unverified Commit 8afff9ae authored by Xiaomeng Zhao's avatar Xiaomeng Zhao Committed by GitHub
Browse files

Merge pull request #1120 from opendatalab/release-0.10.2 

Release 0.10.2
parents 4df1eb74 7fdbb6e5
Show whitespace changes
Inline Side-by-side
Showing with 110 additions and 9128 deletions
magic_pdf/model/magic_model.py
View file @ 8afff9ae
import enum
import json
from magic_pdf.config.model_block_type import ModelBlockTypeEnum
from magic_pdf.config.ocr_content_type import CategoryId, ContentType
from magic_pdf.data.data_reader_writer import (FileBasedDataReader,
FileBasedDataWriter)
from magic_pdf.data.dataset import Dataset
from magic_pdf.libs.boxbase import (_is_in, _is_part_overlap, bbox_distance,
bbox_relative_pos, box_area, calculate_iou,
calculate_overlap_area_in_bbox1_area_ratio,
get_overlap_area)
from magic_pdf.libs.commons import fitz, join_path
from magic_pdf.libs.coordinate_transform import get_scale_ratio
from magic_pdf.libs.local_math import float_gt
from magic_pdf.pre_proc.remove_bbox_overlap import _remove_overlap_between_bbox
......@@ -1048,29 +1044,3 @@ class MagicModel:
def get_model_list(self, page_no):
return self.__model_list[page_no]
if __name__ == '__main__':
drw = FileBasedDataReader(r'D:/project/20231108code-clean')
if 0:
pdf_file_path = r'linshixuqiu\19983-00.pdf'
model_file_path = r'linshixuqiu\19983-00_new.json'
pdf_bytes = drw.read(pdf_file_path)
model_json_txt = drw.read(model_file_path).decode()
model_list = json.loads(model_json_txt)
write_path = r'D:\project\20231108code-clean\linshixuqiu\19983-00'
img_bucket_path = 'imgs'
img_writer = FileBasedDataWriter(join_path(write_path, img_bucket_path))
pdf_docs = fitz.open('pdf', pdf_bytes)
magic_model = MagicModel(model_list, pdf_docs)
if 1:
from magic_pdf.data.dataset import PymuDocDataset
model_list = json.loads(
drw.read('/opt/data/pdf/20240418/j.chroma.2009.03.042.json')
)
pdf_bytes = drw.read('/opt/data/pdf/20240418/j.chroma.2009.03.042.pdf')
magic_model = MagicModel(model_list, PymuDocDataset(pdf_bytes))
for i in range(7):
print(magic_model.get_imgs(i))
magic_pdf/model/sub_modules/ocr/paddleocr/ocr_utils.py
View file @ 8afff9ae
import math
import numpy as np
from loguru import logger
......@@ -214,6 +212,9 @@ def get_ocr_result_list(ocr_res, useful_list):
if len(box_ocr_res) == 2:
p1, p2, p3, p4 = box_ocr_res[0]
text, score = box_ocr_res[1]
# logger.info(f"text: {text}, score: {score}")
if score < 0.6: # 过滤低置信度的结果
continue
else:
p1, p2, p3, p4 = box_ocr_res
text, score = "", 1
......@@ -249,32 +250,6 @@ def get_ocr_result_list(ocr_res, useful_list):
return ocr_result_list
def calculate_angle_degrees(poly):
# 定义对角线的顶点
diagonal1 = (poly[0], poly[2])
diagonal2 = (poly[1], poly[3])
# 计算对角线的斜率
def slope(p1, p2):
return (p2[1] - p1[1]) / (p2[0] - p1[0]) if p2[0] != p1[0] else float('inf')
slope1 = slope(diagonal1[0], diagonal1[1])
slope2 = slope(diagonal2[0], diagonal2[1])
# 计算对角线与x轴的夹角（以弧度为单位）
angle1_radians = math.atan(slope1)
angle2_radians = math.atan(slope2)
# 将弧度转换为角度
angle1_degrees = math.degrees(angle1_radians)
angle2_degrees = math.degrees(angle2_radians)
# 取两条对角线与x轴夹角的平均值
average_angle_degrees = abs((angle1_degrees + angle2_degrees) / 2)
# logger.info(f"average_angle_degrees: {average_angle_degrees}")
return average_angle_degrees
def calculate_is_angle(poly):
p1, p2, p3, p4 = poly
height = ((p4[1] - p1[1]) + (p3[1] - p2[1])) / 2
......
magic_pdf/model/sub_modules/ocr/paddleocr/ppocr_273_mod.py
View file @ 8afff9ae
......@@ -63,7 +63,7 @@ class ModifiedPaddleOCR(PaddleOCR):
if det and rec:
ocr_res = []
for idx, img in enumerate(imgs):
for img in imgs:
img = preprocess_image(img)
dt_boxes, rec_res, _ = self.__call__(img, cls, mfd_res=mfd_res)
if not dt_boxes and not rec_res:
......@@ -75,7 +75,7 @@ class ModifiedPaddleOCR(PaddleOCR):
return ocr_res
elif det and not rec:
ocr_res = []
for idx, img in enumerate(imgs):
for img in imgs:
img = preprocess_image(img)
dt_boxes, elapse = self.text_detector(img)
if dt_boxes is None:
......@@ -96,7 +96,7 @@ class ModifiedPaddleOCR(PaddleOCR):
else:
ocr_res = []
cls_res = []
for idx, img in enumerate(imgs):
for img in imgs:
if not isinstance(img, list):
img = preprocess_image(img)
img = [img]
......
magic_pdf/para/block_continuation_processor.py deleted 100644 → 0
View file @ 4df1eb74
import os
import unicodedata
from magic_pdf.para.commons import *
if sys.version_info[0] >= 3:
sys.stdout.reconfigure(encoding="utf-8") # type: ignore
class BlockContinuationProcessor:
"""
This class is used to process the blocks to detect block continuations.
"""
def __init__(self) -> None:
pass
def __is_similar_font_type(self, font_type1, font_type2, prefix_length_ratio=0.3):
"""
This function checks if the two font types are similar.
Definition of similar font types: the two font types have a common prefix,
and the length of the common prefix is at least a certain ratio of the length of the shorter font type.
Parameters
----------
font_type1 : str
font type 1
font_type2 : str
font type 2
prefix_length_ratio : float
minimum ratio of the common prefix length to the length of the shorter font type
Returns
-------
bool
True if the two font types are similar, False otherwise.
"""
if isinstance(font_type1, list):
font_type1 = font_type1[0] if font_type1 else ""
if isinstance(font_type2, list):
font_type2 = font_type2[0] if font_type2 else ""
if font_type1 == font_type2:
return True
# Find the length of the common prefix
common_prefix_length = len(os.path.commonprefix([font_type1, font_type2]))
# Calculate the minimum prefix length based on the ratio
min_prefix_length = int(min(len(font_type1), len(font_type2)) * prefix_length_ratio)
return common_prefix_length >= min_prefix_length
def __is_same_block_font(self, block1, block2):
"""
This function compares the font of block1 and block2
Parameters
----------
block1 : dict
block1
block2 : dict
block2
Returns
-------
is_same : bool
True if block1 and block2 have the same font, else False
"""
block_1_font_type = safe_get(block1, "block_font_type", "")
block_1_font_size = safe_get(block1, "block_font_size", 0)
block_1_avg_char_width = safe_get(block1, "avg_char_width", 0)
block_2_font_type = safe_get(block2, "block_font_type", "")
block_2_font_size = safe_get(block2, "block_font_size", 0)
block_2_avg_char_width = safe_get(block2, "avg_char_width", 0)
if isinstance(block_1_font_size, list):
block_1_font_size = block_1_font_size[0] if block_1_font_size else 0
if isinstance(block_2_font_size, list):
block_2_font_size = block_2_font_size[0] if block_2_font_size else 0
block_1_text = safe_get(block1, "text", "")
block_2_text = safe_get(block2, "text", "")
if block_1_avg_char_width == 0 or block_2_avg_char_width == 0:
return False
if not block_1_text or not block_2_text:
return False
else:
text_len_ratio = len(block_2_text) / len(block_1_text)
if text_len_ratio < 0.2:
avg_char_width_condition = (
abs(block_1_avg_char_width - block_2_avg_char_width) / min(block_1_avg_char_width, block_2_avg_char_width)
< 0.5
)
else:
avg_char_width_condition = (
abs(block_1_avg_char_width - block_2_avg_char_width) / min(block_1_avg_char_width, block_2_avg_char_width)
< 0.2
)
block_font_size_condtion = abs(block_1_font_size - block_2_font_size) < 1
return (
self.__is_similar_font_type(block_1_font_type, block_2_font_type)
and avg_char_width_condition
and block_font_size_condtion
)
def _is_alphabet_char(self, char):
if (char >= "\u0041" and char <= "\u005a") or (char >= "\u0061" and char <= "\u007a"):
return True
else:
return False
def _is_chinese_char(self, char):
if char >= "\u4e00" and char <= "\u9fa5":
return True
else:
return False
def _is_other_letter_char(self, char):
try:
cat = unicodedata.category(char)
if cat == "Lu" or cat == "Ll":
return not self._is_alphabet_char(char) and not self._is_chinese_char(char)
except TypeError:
print("The input to the function must be a single character.")
return False
def _is_year(self, s: str):
try:
number = int(s)
return 1900 <= number <= 2099
except ValueError:
return False
def __is_para_font_consistent(self, para_1, para_2):
"""
This function compares the font of para1 and para2
Parameters
----------
para1 : dict
para1
para2 : dict
para2
Returns
-------
is_same : bool
True if para1 and para2 have the same font, else False
"""
if para_1 is None or para_2 is None:
return False
para_1_font_type = safe_get(para_1, "para_font_type", "")
para_1_font_size = safe_get(para_1, "para_font_size", 0)
para_1_font_color = safe_get(para_1, "para_font_color", "")
para_2_font_type = safe_get(para_2, "para_font_type", "")
para_2_font_size = safe_get(para_2, "para_font_size", 0)
para_2_font_color = safe_get(para_2, "para_font_color", "")
if isinstance(para_1_font_type, list): # get the most common font type
para_1_font_type = max(set(para_1_font_type), key=para_1_font_type.count)
if isinstance(para_2_font_type, list):
para_2_font_type = max(set(para_2_font_type), key=para_2_font_type.count)
if isinstance(para_1_font_size, list): # compute average font type
para_1_font_size = sum(para_1_font_size) / len(para_1_font_size)
if isinstance(para_2_font_size, list): # compute average font type
para_2_font_size = sum(para_2_font_size) / len(para_2_font_size)
return (
self.__is_similar_font_type(para_1_font_type, para_2_font_type)
and abs(para_1_font_size - para_2_font_size) < 1.5
# and para_font_color1 == para_font_color2
)
def _is_para_puncs_consistent(self, para_1, para_2):
"""
This function determines whether para1 and para2 are originally from the same paragraph by checking the puncs of para1(former) and para2(latter)
Parameters
----------
para1 : dict
para1
para2 : dict
para2
Returns
-------
is_same : bool
True if para1 and para2 are from the same paragraph by using the puncs, else False
"""
para_1_text = safe_get(para_1, "para_text", "").strip()
para_2_text = safe_get(para_2, "para_text", "").strip()
para_1_bboxes = safe_get(para_1, "para_bbox", [])
para_1_font_sizes = safe_get(para_1, "para_font_size", 0)
para_2_bboxes = safe_get(para_2, "para_bbox", [])
para_2_font_sizes = safe_get(para_2, "para_font_size", 0)
# print_yellow(" Features of determine puncs_consistent:")
# print(f" para_1_text: {para_1_text}")
# print(f" para_2_text: {para_2_text}")
# print(f" para_1_bboxes: {para_1_bboxes}")
# print(f" para_2_bboxes: {para_2_bboxes}")
# print(f" para_1_font_sizes: {para_1_font_sizes}")
# print(f" para_2_font_sizes: {para_2_font_sizes}")
if is_nested_list(para_1_bboxes):
x0_1, y0_1, x1_1, y1_1 = para_1_bboxes[-1]
else:
x0_1, y0_1, x1_1, y1_1 = para_1_bboxes
if is_nested_list(para_2_bboxes):
x0_2, y0_2, x1_2, y1_2 = para_2_bboxes[0]
para_2_font_sizes = para_2_font_sizes[0] # type: ignore
else:
x0_2, y0_2, x1_2, y1_2 = para_2_bboxes
right_align_threshold = 0.5 * (para_1_font_sizes + para_2_font_sizes) * 0.8
are_two_paras_right_aligned = abs(x1_1 - x1_2) < right_align_threshold
left_indent_threshold = 0.5 * (para_1_font_sizes + para_2_font_sizes) * 0.8
is_para1_left_indent_than_papa2 = x0_1 - x0_2 > left_indent_threshold
is_para2_left_indent_than_papa1 = x0_2 - x0_1 > left_indent_threshold
# Check if either para_text1 or para_text2 is empty
if not para_1_text or not para_2_text:
return False
# Define the end puncs for a sentence to end and hyphen
end_puncs = [".", "?", "!", "。", "？", "！", "…"]
hyphen = ["-", "—"]
# Check if para_text1 ends with either hyphen or non-end punctuation or spaces
para_1_end_with_hyphen = para_1_text and para_1_text[-1] in hyphen
para_1_end_with_end_punc = para_1_text and para_1_text[-1] in end_puncs
para_1_end_with_space = para_1_text and para_1_text[-1] == " "
para_1_not_end_with_end_punc = para_1_text and para_1_text[-1] not in end_puncs
# print_yellow(f" para_1_end_with_hyphen: {para_1_end_with_hyphen}")
# print_yellow(f" para_1_end_with_end_punc: {para_1_end_with_end_punc}")
# print_yellow(f" para_1_not_end_with_end_punc: {para_1_not_end_with_end_punc}")
# print_yellow(f" para_1_end_with_space: {para_1_end_with_space}")
if para_1_end_with_hyphen: # If para_text1 ends with hyphen
# print_red(f"para_1 is end with hyphen.")
para_2_is_consistent = para_2_text and (
para_2_text[0] in hyphen
or (self._is_alphabet_char(para_2_text[0]) and para_2_text[0].islower())
or (self._is_chinese_char(para_2_text[0]))
or (self._is_other_letter_char(para_2_text[0]))
)
if para_2_is_consistent:
# print(f"para_2 is consistent.\n")
return True
else:
# print(f"para_2 is not consistent.\n")
pass
elif para_1_end_with_end_punc: # If para_text1 ends with ending punctuations
# print_red(f"para_1 is end with end_punc.")
para_2_is_consistent = (
para_2_text
and (
para_2_text[0] == " "
or (self._is_alphabet_char(para_2_text[0]) and para_2_text[0].isupper())
or (self._is_chinese_char(para_2_text[0]))
or (self._is_other_letter_char(para_2_text[0]))
)
and not is_para2_left_indent_than_papa1
)
if para_2_is_consistent:
# print(f"para_2 is consistent.\n")
return True
else:
# print(f"para_2 is not consistent.\n")
pass
elif para_1_not_end_with_end_punc: # If para_text1 is not end with ending punctuations
# print_red(f"para_1 is NOT end with end_punc.")
para_2_is_consistent = para_2_text and (
para_2_text[0] == " "
or (self._is_alphabet_char(para_2_text[0]) and para_2_text[0].islower())
or (self._is_alphabet_char(para_2_text[0]))
or (self._is_year(para_2_text[0:4]))
or (are_two_paras_right_aligned or is_para1_left_indent_than_papa2)
or (self._is_chinese_char(para_2_text[0]))
or (self._is_other_letter_char(para_2_text[0]))
)
if para_2_is_consistent:
# print(f"para_2 is consistent.\n")
return True
else:
# print(f"para_2 is not consistent.\n")
pass
elif para_1_end_with_space: # If para_text1 ends with space
# print_red(f"para_1 is end with space.")
para_2_is_consistent = para_2_text and (
para_2_text[0] == " "
or (self._is_alphabet_char(para_2_text[0]) and para_2_text[0].islower())
or (self._is_chinese_char(para_2_text[0]))
or (self._is_other_letter_char(para_2_text[0]))
)
if para_2_is_consistent:
# print(f"para_2 is consistent.\n")
return True
else:
pass
# print(f"para_2 is not consistent.\n")
return False
def _is_block_consistent(self, block1, block2):
"""
This function determines whether block1 and block2 are originally from the same block
Parameters
----------
block1 : dict
block1s
block2 : dict
block2
Returns
-------
is_same : bool
True if block1 and block2 are from the same block, else False
"""
return self.__is_same_block_font(block1, block2)
def _is_para_continued(self, para1, para2):
"""
This function determines whether para1 and para2 are originally from the same paragraph
Parameters
----------
para1 : dict
para1
para2 : dict
para2
Returns
-------
is_same : bool
True if para1 and para2 are from the same paragraph, else False
"""
is_para_font_consistent = self.__is_para_font_consistent(para1, para2)
is_para_puncs_consistent = self._is_para_puncs_consistent(para1, para2)
return is_para_font_consistent and is_para_puncs_consistent
def _are_boundaries_of_block_consistent(self, block1, block2):
"""
This function checks if the boundaries of block1 and block2 are consistent
Parameters
----------
block1 : dict
block1
block2 : dict
block2
Returns
-------
is_consistent : bool
True if the boundaries of block1 and block2 are consistent, else False
"""
last_line_of_block1 = block1["lines"][-1]
first_line_of_block2 = block2["lines"][0]
spans_of_last_line_of_block1 = last_line_of_block1["spans"]
spans_of_first_line_of_block2 = first_line_of_block2["spans"]
font_type_of_last_line_of_block1 = spans_of_last_line_of_block1[0]["font"].lower()
font_size_of_last_line_of_block1 = spans_of_last_line_of_block1[0]["size"]
font_color_of_last_line_of_block1 = spans_of_last_line_of_block1[0]["color"]
font_flags_of_last_line_of_block1 = spans_of_last_line_of_block1[0]["flags"]
font_type_of_first_line_of_block2 = spans_of_first_line_of_block2[0]["font"].lower()
font_size_of_first_line_of_block2 = spans_of_first_line_of_block2[0]["size"]
font_color_of_first_line_of_block2 = spans_of_first_line_of_block2[0]["color"]
font_flags_of_first_line_of_block2 = spans_of_first_line_of_block2[0]["flags"]
return (
self.__is_similar_font_type(font_type_of_last_line_of_block1, font_type_of_first_line_of_block2)
and abs(font_size_of_last_line_of_block1 - font_size_of_first_line_of_block2) < 1
# and font_color_of_last_line_of_block1 == font_color_of_first_line_of_block2
and font_flags_of_last_line_of_block1 == font_flags_of_first_line_of_block2
)
def _get_last_paragraph(self, block):
"""
Retrieves the last paragraph from a block.
Parameters
----------
block : dict
The block from which to retrieve the paragraph.
Returns
-------
dict
The last paragraph of the block.
"""
if block["paras"]:
last_para_key = list(block["paras"].keys())[-1]
return block["paras"][last_para_key]
else:
return None
def _get_first_paragraph(self, block):
"""
Retrieves the first paragraph from a block.
Parameters
----------
block : dict
The block from which to retrieve the paragraph.
Returns
-------
dict
The first paragraph of the block.
"""
if block["paras"]:
first_para_key = list(block["paras"].keys())[0]
return block["paras"][first_para_key]
else:
return None
def should_merge_next_para(self, curr_para, next_para):
if self._is_para_continued(curr_para, next_para):
return True
else:
return False
def batch_tag_paras(self, pdf_dict):
the_last_page_id = len(pdf_dict) - 1
for curr_page_idx, (curr_page_id, curr_page_content) in enumerate(pdf_dict.items()):
if curr_page_id.startswith("page_") and curr_page_content.get("para_blocks", []):
para_blocks_of_curr_page = curr_page_content["para_blocks"]
next_page_idx = curr_page_idx + 1
next_page_id = f"page_{next_page_idx}"
next_page_content = pdf_dict.get(next_page_id, {})
for i, current_block in enumerate(para_blocks_of_curr_page):
for para_id, curr_para in current_block["paras"].items():
curr_para["curr_para_location"] = [
curr_page_idx,
current_block["block_id"],
int(para_id.split("_")[-1]),
]
curr_para["next_para_location"] = None # 默认设置为None
curr_para["merge_next_para"] = False # 默认设置为False
next_block = para_blocks_of_curr_page[i + 1] if i < len(para_blocks_of_curr_page) - 1 else None
if next_block:
curr_block_last_para_key = list(current_block["paras"].keys())[-1]
curr_blk_last_para = current_block["paras"][curr_block_last_para_key]
next_block_first_para_key = list(next_block["paras"].keys())[0]
next_blk_first_para = next_block["paras"][next_block_first_para_key]
if self.should_merge_next_para(curr_blk_last_para, next_blk_first_para):
curr_blk_last_para["next_para_location"] = [
curr_page_idx,
next_block["block_id"],
int(next_block_first_para_key.split("_")[-1]),
]
curr_blk_last_para["merge_next_para"] = True
else:
# Handle the case where the next block is in a different page
curr_block_last_para_key = list(current_block["paras"].keys())[-1]
curr_blk_last_para = current_block["paras"][curr_block_last_para_key]
while not next_page_content.get("para_blocks", []) and next_page_idx <= the_last_page_id:
next_page_idx += 1
next_page_id = f"page_{next_page_idx}"
next_page_content = pdf_dict.get(next_page_id, {})
if next_page_content.get("para_blocks", []):
next_blk_first_para_key = list(next_page_content["para_blocks"][0]["paras"].keys())[0]
next_blk_first_para = next_page_content["para_blocks"][0]["paras"][next_blk_first_para_key]
if self.should_merge_next_para(curr_blk_last_para, next_blk_first_para):
curr_blk_last_para["next_para_location"] = [
next_page_idx,
next_page_content["para_blocks"][0]["block_id"],
int(next_blk_first_para_key.split("_")[-1]),
]
curr_blk_last_para["merge_next_para"] = True
return pdf_dict
def find_block_by_id(self, para_blocks, block_id):
for block in para_blocks:
if block.get("block_id") == block_id:
return block
return None
def batch_merge_paras(self, pdf_dict):
for page_id, page_content in pdf_dict.items():
if page_id.startswith("page_") and page_content.get("para_blocks", []):
para_blocks_of_page = page_content["para_blocks"]
for i in range(len(para_blocks_of_page)):
current_block = para_blocks_of_page[i]
paras = current_block["paras"]
for para_id, curr_para in list(paras.items()):
# 跳过标题段落
if curr_para.get("is_para_title"):
continue
while curr_para.get("merge_next_para"):
next_para_location = curr_para.get("next_para_location")
if not next_para_location:
break
next_page_idx, next_block_id, next_para_id = next_para_location
next_page_id = f"page_{next_page_idx}"
next_page_content = pdf_dict.get(next_page_id)
if not next_page_content:
break
next_block = self.find_block_by_id(next_page_content.get("para_blocks", []), next_block_id)
if not next_block:
break
next_para = next_block["paras"].get(f"para_{next_para_id}")
if not next_para or next_para.get("is_para_title"):
break
# 合并段落文本
curr_para_text = curr_para.get("para_text", "")
next_para_text = next_para.get("para_text", "")
curr_para["para_text"] = curr_para_text + " " + next_para_text
# 更新 next_para_location
curr_para["next_para_location"] = next_para.get("next_para_location")
# 将下一个段落文本置为空，表示已被合并
next_para["para_text"] = ""
# 更新 merge_next_para 标记
curr_para["merge_next_para"] = next_para.get("merge_next_para", False)
return pdf_dict
magic_pdf/para/block_termination_processor.py deleted 100644 → 0
View file @ 4df1eb74
from magic_pdf.para.commons import *
if sys.version_info[0] >= 3:
sys.stdout.reconfigure(encoding="utf-8") # type: ignore
class BlockTerminationProcessor:
def __init__(self) -> None:
pass
def _is_consistent_lines(
self,
curr_line,
prev_line,
next_line,
consistent_direction, # 0 for prev, 1 for next, 2 for both
):
"""
This function checks if the line is consistent with its neighbors
Parameters
----------
curr_line : dict
current line
prev_line : dict
previous line
next_line : dict
next line
consistent_direction : int
0 for prev, 1 for next, 2 for both
Returns
-------
bool
True if the line is consistent with its neighbors, False otherwise.
"""
curr_line_font_size = curr_line["spans"][0]["size"]
curr_line_font_type = curr_line["spans"][0]["font"].lower()
if consistent_direction == 0:
if prev_line:
prev_line_font_size = prev_line["spans"][0]["size"]
prev_line_font_type = prev_line["spans"][0]["font"].lower()
return curr_line_font_size == prev_line_font_size and curr_line_font_type == prev_line_font_type
else:
return False
elif consistent_direction == 1:
if next_line:
next_line_font_size = next_line["spans"][0]["size"]
next_line_font_type = next_line["spans"][0]["font"].lower()
return curr_line_font_size == next_line_font_size and curr_line_font_type == next_line_font_type
else:
return False
elif consistent_direction == 2:
if prev_line and next_line:
prev_line_font_size = prev_line["spans"][0]["size"]
prev_line_font_type = prev_line["spans"][0]["font"].lower()
next_line_font_size = next_line["spans"][0]["size"]
next_line_font_type = next_line["spans"][0]["font"].lower()
return (curr_line_font_size == prev_line_font_size and curr_line_font_type == prev_line_font_type) and (
curr_line_font_size == next_line_font_size and curr_line_font_type == next_line_font_type
)
else:
return False
else:
return False
def _is_regular_line(self, curr_line_bbox, prev_line_bbox, next_line_bbox, avg_char_width, X0, X1, avg_line_height):
"""
This function checks if the line is a regular line
Parameters
----------
curr_line_bbox : list
bbox of the current line
prev_line_bbox : list
bbox of the previous line
next_line_bbox : list
bbox of the next line
avg_char_width : float
average of char widths
X0 : float
median of x0 values, which represents the left average boundary of the page
X1 : float
median of x1 values, which represents the right average boundary of the page
avg_line_height : float
average of line heights
Returns
-------
bool
True if the line is a regular line, False otherwise.
"""
horizontal_ratio = 0.5
vertical_ratio = 0.5
horizontal_thres = horizontal_ratio * avg_char_width
vertical_thres = vertical_ratio * avg_line_height
x0, y0, x1, y1 = curr_line_bbox
x0_near_X0 = abs(x0 - X0) < horizontal_thres
x1_near_X1 = abs(x1 - X1) < horizontal_thres
prev_line_is_end_of_para = prev_line_bbox and (abs(prev_line_bbox[2] - X1) > avg_char_width)
sufficient_spacing_above = False
if prev_line_bbox:
vertical_spacing_above = y1 - prev_line_bbox[3]
sufficient_spacing_above = vertical_spacing_above > vertical_thres
sufficient_spacing_below = False
if next_line_bbox:
vertical_spacing_below = next_line_bbox[1] - y0
sufficient_spacing_below = vertical_spacing_below > vertical_thres
return (
(sufficient_spacing_above or sufficient_spacing_below)
or (not x0_near_X0 and not x1_near_X1)
or prev_line_is_end_of_para
)
def _is_possible_start_of_para(self, curr_line, prev_line, next_line, X0, X1, avg_char_width, avg_font_size):
"""
This function checks if the line is a possible start of a paragraph
Parameters
----------
curr_line : dict
current line
prev_line : dict
previous line
next_line : dict
next line
X0 : float
median of x0 values, which represents the left average boundary of the page
X1 : float
median of x1 values, which represents the right average boundary of the page
avg_char_width : float
average of char widths
avg_line_height : float
average of line heights
Returns
-------
bool
True if the line is a possible start of a paragraph, False otherwise.
"""
start_confidence = 0.5 # Initial confidence of the line being a start of a paragraph
decision_path = [] # Record the decision path
curr_line_bbox = curr_line["bbox"]
prev_line_bbox = prev_line["bbox"] if prev_line else None
next_line_bbox = next_line["bbox"] if next_line else None
indent_ratio = 1
vertical_ratio = 1.5
vertical_thres = vertical_ratio * avg_font_size
left_horizontal_ratio = 0.5
left_horizontal_thres = left_horizontal_ratio * avg_char_width
right_horizontal_ratio = 2.5
right_horizontal_thres = right_horizontal_ratio * avg_char_width
x0, y0, x1, y1 = curr_line_bbox
indent_condition = x0 > X0 + indent_ratio * avg_char_width
if indent_condition:
start_confidence += 0.2
decision_path.append("indent_condition_met")
x0_near_X0 = abs(x0 - X0) < left_horizontal_thres
if x0_near_X0:
start_confidence += 0.1
decision_path.append("x0_near_X0")
x1_near_X1 = abs(x1 - X1) < right_horizontal_thres
if x1_near_X1:
start_confidence += 0.1
decision_path.append("x1_near_X1")
if prev_line is None:
prev_line_is_end_of_para = True
start_confidence += 0.2
decision_path.append("no_prev_line")
else:
prev_line_is_end_of_para, _, _ = self._is_possible_end_of_para(prev_line, next_line, X0, X1, avg_char_width)
if prev_line_is_end_of_para:
start_confidence += 0.1
decision_path.append("prev_line_is_end_of_para")
sufficient_spacing_above = False
if prev_line_bbox:
vertical_spacing_above = y1 - prev_line_bbox[3]
sufficient_spacing_above = vertical_spacing_above > vertical_thres
if sufficient_spacing_above:
start_confidence += 0.2
decision_path.append("sufficient_spacing_above")
sufficient_spacing_below = False
if next_line_bbox:
vertical_spacing_below = next_line_bbox[1] - y0
sufficient_spacing_below = vertical_spacing_below > vertical_thres
if sufficient_spacing_below:
start_confidence += 0.2
decision_path.append("sufficient_spacing_below")
is_regular_line = self._is_regular_line(
curr_line_bbox, prev_line_bbox, next_line_bbox, avg_char_width, X0, X1, avg_font_size
)
if is_regular_line:
start_confidence += 0.1
decision_path.append("is_regular_line")
is_start_of_para = (
(sufficient_spacing_above or sufficient_spacing_below)
or (indent_condition)
or (not indent_condition and x0_near_X0 and x1_near_X1 and not is_regular_line)
or prev_line_is_end_of_para
)
return (is_start_of_para, start_confidence, decision_path)
def _is_possible_end_of_para(self, curr_line, next_line, X0, X1, avg_char_width):
"""
This function checks if the line is a possible end of a paragraph
Parameters
----------
curr_line : dict
current line
next_line : dict
next line
X0 : float
median of x0 values, which represents the left average boundary of the page
X1 : float
median of x1 values, which represents the right average boundary of the page
avg_char_width : float
average of char widths
Returns
-------
bool
True if the line is a possible end of a paragraph, False otherwise.
"""
end_confidence = 0.5 # Initial confidence of the line being a end of a paragraph
decision_path = [] # Record the decision path
curr_line_bbox = curr_line["bbox"]
next_line_bbox = next_line["bbox"] if next_line else None
left_horizontal_ratio = 0.5
right_horizontal_ratio = 0.5
x0, _, x1, y1 = curr_line_bbox
next_x0, next_y0, _, _ = next_line_bbox if next_line_bbox else (0, 0, 0, 0)
x0_near_X0 = abs(x0 - X0) < left_horizontal_ratio * avg_char_width
if x0_near_X0:
end_confidence += 0.1
decision_path.append("x0_near_X0")
x1_smaller_than_X1 = x1 < X1 - right_horizontal_ratio * avg_char_width
if x1_smaller_than_X1:
end_confidence += 0.1
decision_path.append("x1_smaller_than_X1")
next_line_is_start_of_para = (
next_line_bbox
and (next_x0 > X0 + left_horizontal_ratio * avg_char_width)
and (not is_line_left_aligned_from_neighbors(curr_line_bbox, None, next_line_bbox, avg_char_width, direction=1))
)
if next_line_is_start_of_para:
end_confidence += 0.2
decision_path.append("next_line_is_start_of_para")
is_line_left_aligned_from_neighbors_bool = is_line_left_aligned_from_neighbors(
curr_line_bbox, None, next_line_bbox, avg_char_width
)
if is_line_left_aligned_from_neighbors_bool:
end_confidence += 0.1
decision_path.append("line_is_left_aligned_from_neighbors")
is_line_right_aligned_from_neighbors_bool = is_line_right_aligned_from_neighbors(
curr_line_bbox, None, next_line_bbox, avg_char_width
)
if not is_line_right_aligned_from_neighbors_bool:
end_confidence += 0.1
decision_path.append("line_is_not_right_aligned_from_neighbors")
is_end_of_para = end_with_punctuation(curr_line["text"]) and (
(x0_near_X0 and x1_smaller_than_X1)
or (is_line_left_aligned_from_neighbors_bool and not is_line_right_aligned_from_neighbors_bool)
)
return (is_end_of_para, end_confidence, decision_path)
def _cut_paras_per_block(
self,
block,
):
"""
Processes a raw block from PyMuPDF and returns the processed block.
Parameters
----------
raw_block : dict
A raw block from pymupdf.
Returns
-------
processed_block : dict
"""
def _construct_para(lines, is_block_title, para_title_level):
"""
Construct a paragraph from given lines.
"""
font_sizes = [span["size"] for line in lines for span in line["spans"]]
avg_font_size = sum(font_sizes) / len(font_sizes) if font_sizes else 0
font_colors = [span["color"] for line in lines for span in line["spans"]]
most_common_font_color = max(set(font_colors), key=font_colors.count) if font_colors else None
# font_types = [span["font"] for line in lines for span in line["spans"]]
# most_common_font_type = max(set(font_types), key=font_types.count) if font_types else None
font_type_lengths = {}
for line in lines:
for span in line["spans"]:
font_type = span["font"]
bbox_width = span["bbox"][2] - span["bbox"][0]
if font_type in font_type_lengths:
font_type_lengths[font_type] += bbox_width
else:
font_type_lengths[font_type] = bbox_width
# get the font type with the longest bbox width
most_common_font_type = max(font_type_lengths, key=font_type_lengths.get) if font_type_lengths else None # type: ignore
para_bbox = calculate_para_bbox(lines)
para_text = " ".join(line["text"] for line in lines)
return {
"para_bbox": para_bbox,
"para_text": para_text,
"para_font_type": most_common_font_type,
"para_font_size": avg_font_size,
"para_font_color": most_common_font_color,
"is_para_title": is_block_title,
"para_title_level": para_title_level,
}
block_bbox = block["bbox"]
block_text = block["text"]
block_lines = block["lines"]
X0 = safe_get(block, "X0", 0)
X1 = safe_get(block, "X1", 0)
avg_char_width = safe_get(block, "avg_char_width", 0)
avg_char_height = safe_get(block, "avg_char_height", 0)
avg_font_size = safe_get(block, "avg_font_size", 0)
is_block_title = safe_get(block, "is_block_title", False)
para_title_level = safe_get(block, "block_title_level", 0)
# Segment into paragraphs
para_ranges = []
in_paragraph = False
start_idx_of_para = None
# Create the processed paragraphs
processed_paras = {}
para_bboxes = []
end_idx_of_para = 0
for line_index, line in enumerate(block_lines):
curr_line = line
prev_line = block_lines[line_index - 1] if line_index > 0 else None
next_line = block_lines[line_index + 1] if line_index < len(block_lines) - 1 else None
"""
Start processing paragraphs.
"""
# Check if the line is the start of a paragraph
is_start_of_para, start_confidence, decision_path = self._is_possible_start_of_para(
curr_line, prev_line, next_line, X0, X1, avg_char_width, avg_font_size
)
if not in_paragraph and is_start_of_para:
in_paragraph = True
start_idx_of_para = line_index
# print_green(">>> Start of a paragraph")
# print(" curr_line_text: ", curr_line["text"])
# print(" start_confidence: ", start_confidence)
# print(" decision_path: ", decision_path)
# Check if the line is the end of a paragraph
is_end_of_para, end_confidence, decision_path = self._is_possible_end_of_para(
curr_line, next_line, X0, X1, avg_char_width
)
if in_paragraph and (is_end_of_para or not next_line):
para_ranges.append((start_idx_of_para, line_index))
start_idx_of_para = None
in_paragraph = False
# print_red(">>> End of a paragraph")
# print(" curr_line_text: ", curr_line["text"])
# print(" end_confidence: ", end_confidence)
# print(" decision_path: ", decision_path)
# Add the last paragraph if it is not added
if in_paragraph and start_idx_of_para is not None:
para_ranges.append((start_idx_of_para, len(block_lines) - 1))
# Process the matched paragraphs
for para_index, (start_idx, end_idx) in enumerate(para_ranges):
matched_lines = block_lines[start_idx : end_idx + 1]
para_properties = _construct_para(matched_lines, is_block_title, para_title_level)
para_key = f"para_{len(processed_paras)}"
processed_paras[para_key] = para_properties
para_bboxes.append(para_properties["para_bbox"])
end_idx_of_para = end_idx + 1
# Deal with the remaining lines
if end_idx_of_para < len(block_lines):
unmatched_lines = block_lines[end_idx_of_para:]
unmatched_properties = _construct_para(unmatched_lines, is_block_title, para_title_level)
unmatched_key = f"para_{len(processed_paras)}"
processed_paras[unmatched_key] = unmatched_properties
para_bboxes.append(unmatched_properties["para_bbox"])
block["paras"] = processed_paras
return block
def batch_process_blocks(self, pdf_dict):
"""
Parses the blocks of all pages.
Parameters
----------
pdf_dict : dict
PDF dictionary.
filter_blocks : list
List of bounding boxes to filter.
Returns
-------
result_dict : dict
Result dictionary.
"""
num_paras = 0
for page_id, page in pdf_dict.items():
if page_id.startswith("page_"):
para_blocks = []
if "para_blocks" in page.keys():
input_blocks = page["para_blocks"]
for input_block in input_blocks:
new_block = self._cut_paras_per_block(input_block)
para_blocks.append(new_block)
num_paras += len(new_block["paras"])
page["para_blocks"] = para_blocks
pdf_dict["statistics"]["num_paras"] = num_paras
return pdf_dict
magic_pdf/para/commons.py deleted 100644 → 0
View file @ 4df1eb74
import sys
from magic_pdf.libs.commons import fitz
from termcolor import cprint
if sys.version_info[0] >= 3:
sys.stdout.reconfigure(encoding="utf-8") # type: ignore
def open_pdf(pdf_path):
try:
pdf_document = fitz.open(pdf_path) # type: ignore
return pdf_document
except Exception as e:
print(f"无法打开PDF文件：{pdf_path}。原因是：{e}")
raise e
def print_green_on_red(text):
cprint(text, "green", "on_red", attrs=["bold"], end="\n\n")
def print_green(text):
print()
cprint(text, "green", attrs=["bold"], end="\n\n")
def print_red(text):
print()
cprint(text, "red", attrs=["bold"], end="\n\n")
def print_yellow(text):
print()
cprint(text, "yellow", attrs=["bold"], end="\n\n")
def safe_get(dict_obj, key, default):
val = dict_obj.get(key)
if val is None:
return default
else:
return val
def is_bbox_overlap(bbox1, bbox2):
"""
This function checks if bbox1 and bbox2 overlap or not
Parameters
----------
bbox1 : list
bbox1
bbox2 : list
bbox2
Returns
-------
bool
True if bbox1 and bbox2 overlap, else False
"""
x0_1, y0_1, x1_1, y1_1 = bbox1
x0_2, y0_2, x1_2, y1_2 = bbox2
if x0_1 > x1_2 or x0_2 > x1_1:
return False
if y0_1 > y1_2 or y0_2 > y1_1:
return False
return True
def is_in_bbox(bbox1, bbox2):
"""
This function checks if bbox1 is in bbox2
Parameters
----------
bbox1 : list
bbox1
bbox2 : list
bbox2
Returns
-------
bool
True if bbox1 is in bbox2, else False
"""
x0_1, y0_1, x1_1, y1_1 = bbox1
x0_2, y0_2, x1_2, y1_2 = bbox2
if x0_1 >= x0_2 and y0_1 >= y0_2 and x1_1 <= x1_2 and y1_1 <= y1_2:
return True
else:
return False
def calculate_para_bbox(lines):
"""
This function calculates the minimum bbox of the paragraph
Parameters
----------
lines : list
lines
Returns
-------
para_bbox : list
bbox of the paragraph
"""
x0 = min(line["bbox"][0] for line in lines)
y0 = min(line["bbox"][1] for line in lines)
x1 = max(line["bbox"][2] for line in lines)
y1 = max(line["bbox"][3] for line in lines)
return [x0, y0, x1, y1]
def is_line_right_aligned_from_neighbors(curr_line_bbox, prev_line_bbox, next_line_bbox, avg_char_width, direction=2):
"""
This function checks if the line is right aligned from its neighbors
Parameters
----------
curr_line_bbox : list
bbox of the current line
prev_line_bbox : list
bbox of the previous line
next_line_bbox : list
bbox of the next line
avg_char_width : float
average of char widths
direction : int
0 for prev, 1 for next, 2 for both
Returns
-------
bool
True if the line is right aligned from its neighbors, False otherwise.
"""
horizontal_ratio = 0.5
horizontal_thres = horizontal_ratio * avg_char_width
_, _, x1, _ = curr_line_bbox
_, _, prev_x1, _ = prev_line_bbox if prev_line_bbox else (0, 0, 0, 0)
_, _, next_x1, _ = next_line_bbox if next_line_bbox else (0, 0, 0, 0)
if direction == 0:
return abs(x1 - prev_x1) < horizontal_thres
elif direction == 1:
return abs(x1 - next_x1) < horizontal_thres
elif direction == 2:
return abs(x1 - prev_x1) < horizontal_thres and abs(x1 - next_x1) < horizontal_thres
else:
return False
def is_line_left_aligned_from_neighbors(curr_line_bbox, prev_line_bbox, next_line_bbox, avg_char_width, direction=2):
"""
This function checks if the line is left aligned from its neighbors
Parameters
----------
curr_line_bbox : list
bbox of the current line
prev_line_bbox : list
bbox of the previous line
next_line_bbox : list
bbox of the next line
avg_char_width : float
average of char widths
direction : int
0 for prev, 1 for next, 2 for both
Returns
-------
bool
True if the line is left aligned from its neighbors, False otherwise.
"""
horizontal_ratio = 0.5
horizontal_thres = horizontal_ratio * avg_char_width
x0, _, _, _ = curr_line_bbox
prev_x0, _, _, _ = prev_line_bbox if prev_line_bbox else (0, 0, 0, 0)
next_x0, _, _, _ = next_line_bbox if next_line_bbox else (0, 0, 0, 0)
if direction == 0:
return abs(x0 - prev_x0) < horizontal_thres
elif direction == 1:
return abs(x0 - next_x0) < horizontal_thres
elif direction == 2:
return abs(x0 - prev_x0) < horizontal_thres and abs(x0 - next_x0) < horizontal_thres
else:
return False
def end_with_punctuation(line_text):
"""
This function checks if the line ends with punctuation marks
"""
english_end_puncs = [".", "?", "!"]
chinese_end_puncs = ["。", "？", "！"]
end_puncs = english_end_puncs + chinese_end_puncs
last_non_space_char = None
for ch in line_text[::-1]:
if not ch.isspace():
last_non_space_char = ch
break
if last_non_space_char is None:
return False
return last_non_space_char in end_puncs
def is_nested_list(lst):
if isinstance(lst, list):
return any(isinstance(sub, list) for sub in lst)
return False
magic_pdf/para/denoise.py deleted 100644 → 0
View file @ 4df1eb74
import math
from collections import defaultdict
from magic_pdf.para.commons import *
if sys.version_info[0] >= 3:
sys.stdout.reconfigure(encoding="utf-8") # type: ignore
class HeaderFooterProcessor:
def __init__(self) -> None:
pass
def get_most_common_bboxes(self, bboxes, page_height, position="top", threshold=0.25, num_bboxes=3, min_frequency=2):
"""
This function gets the most common bboxes from the bboxes
Parameters
----------
bboxes : list
bboxes
page_height : float
height of the page
position : str, optional
"top" or "bottom", by default "top"
threshold : float, optional
threshold, by default 0.25
num_bboxes : int, optional
number of bboxes to return, by default 3
min_frequency : int, optional
minimum frequency of the bbox, by default 2
Returns
-------
common_bboxes : list
common bboxes
"""
# Filter bbox by position
if position == "top":
filtered_bboxes = [bbox for bbox in bboxes if bbox[1] < page_height * threshold]
else:
filtered_bboxes = [bbox for bbox in bboxes if bbox[3] > page_height * (1 - threshold)]
# Find the most common bbox
bbox_count = defaultdict(int)
for bbox in filtered_bboxes:
bbox_count[tuple(bbox)] += 1
# Get the most frequently occurring bbox, but only consider it when the frequency exceeds min_frequency
common_bboxes = [
bbox for bbox, count in sorted(bbox_count.items(), key=lambda item: item[1], reverse=True) if count >= min_frequency
][:num_bboxes]
return common_bboxes
def detect_footer_header(self, result_dict, similarity_threshold=0.5):
"""
This function detects the header and footer of the document.
Parameters
----------
result_dict : dict
result dictionary
Returns
-------
result_dict : dict
result dictionary
"""
def compare_bbox_with_list(bbox, bbox_list, tolerance=1):
return any(all(abs(a - b) < tolerance for a, b in zip(bbox, common_bbox)) for common_bbox in bbox_list)
def is_single_line_block(block):
# Determine based on the width and height of the block
block_width = block["X1"] - block["X0"]
block_height = block["bbox"][3] - block["bbox"][1]
# If the height of the block is close to the average character height and the width is large, it is considered a single line
return block_height <= block["avg_char_height"] * 3 and block_width > block["avg_char_width"] * 3
# Traverse all blocks in the document
single_preproc_blocks = 0
total_blocks = 0
single_preproc_blocks = 0
for page_id, blocks in result_dict.items():
if page_id.startswith("page_"):
for block_key, block in blocks.items():
if block_key.startswith("block_"):
total_blocks += 1
if is_single_line_block(block):
single_preproc_blocks += 1
# If there are no blocks, skip the header and footer detection
if total_blocks == 0:
print("No blocks found. Skipping header/footer detection.")
return result_dict
# If most of the blocks are single-line, skip the header and footer detection
if single_preproc_blocks / total_blocks > 0.5: # 50% of the blocks are single-line
return result_dict
# Collect the bounding boxes of all blocks
all_bboxes = []
all_texts = []
for page_id, blocks in result_dict.items():
if page_id.startswith("page_"):
for block_key, block in blocks.items():
if block_key.startswith("block_"):
all_bboxes.append(block["bbox"])
# Get the height of the page
page_height = max(bbox[3] for bbox in all_bboxes)
# Get the most common bbox lists for headers and footers
common_header_bboxes = self.get_most_common_bboxes(all_bboxes, page_height, position="top") if all_bboxes else []
common_footer_bboxes = self.get_most_common_bboxes(all_bboxes, page_height, position="bottom") if all_bboxes else []
# Detect and mark headers and footers
for page_id, blocks in result_dict.items():
if page_id.startswith("page_"):
for block_key, block in blocks.items():
if block_key.startswith("block_"):
bbox = block["bbox"]
text = block["text"]
is_header = compare_bbox_with_list(bbox, common_header_bboxes)
is_footer = compare_bbox_with_list(bbox, common_footer_bboxes)
block["is_header"] = int(is_header)
block["is_footer"] = int(is_footer)
return result_dict
class NonHorizontalTextProcessor:
def __init__(self) -> None:
pass
def detect_non_horizontal_texts(self, result_dict):
"""
This function detects watermarks and vertical margin notes in the document.
Watermarks are identified by finding blocks with the same coordinates and frequently occurring identical texts across multiple pages.
If these conditions are met, the blocks are highly likely to be watermarks, as opposed to headers or footers, which can change from page to page.
If the direction of these blocks is not horizontal, they are definitely considered to be watermarks.
Vertical margin notes are identified by finding blocks with the same coordinates and frequently occurring identical texts across multiple pages.
If these conditions are met, the blocks are highly likely to be vertical margin notes, which typically appear on the left and right sides of the page.
If the direction of these blocks is vertical, they are definitely considered to be vertical margin notes.
Parameters
----------
result_dict : dict
The result dictionary.
Returns
-------
result_dict : dict
The updated result dictionary.
"""
# Dictionary to store information about potential watermarks
potential_watermarks = {}
potential_margin_notes = {}
for page_id, page_content in result_dict.items():
if page_id.startswith("page_"):
for block_id, block_data in page_content.items():
if block_id.startswith("block_"):
if "dir" in block_data:
coordinates_text = (block_data["bbox"], block_data["text"]) # Tuple of coordinates and text
angle = math.atan2(block_data["dir"][1], block_data["dir"][0])
angle = abs(math.degrees(angle))
if angle > 5 and angle < 85: # Check if direction is watermarks
if coordinates_text in potential_watermarks:
potential_watermarks[coordinates_text] += 1
else:
potential_watermarks[coordinates_text] = 1
if angle > 85 and angle < 105: # Check if direction is vertical
if coordinates_text in potential_margin_notes:
potential_margin_notes[coordinates_text] += 1 # Increment count
else:
potential_margin_notes[coordinates_text] = 1 # Initialize count
# Identify watermarks by finding entries with counts higher than a threshold (e.g., appearing on more than half of the pages)
watermark_threshold = len(result_dict) // 2
watermarks = {k: v for k, v in potential_watermarks.items() if v > watermark_threshold}
# Identify margin notes by finding entries with counts higher than a threshold (e.g., appearing on more than half of the pages)
margin_note_threshold = len(result_dict) // 2
margin_notes = {k: v for k, v in potential_margin_notes.items() if v > margin_note_threshold}
# Add watermark information to the result dictionary
for page_id, blocks in result_dict.items():
if page_id.startswith("page_"):
for block_id, block_data in blocks.items():
coordinates_text = (block_data["bbox"], block_data["text"])
if coordinates_text in watermarks:
block_data["is_watermark"] = 1
else:
block_data["is_watermark"] = 0
if coordinates_text in margin_notes:
block_data["is_vertical_margin_note"] = 1
else:
block_data["is_vertical_margin_note"] = 0
return result_dict
class NoiseRemover:
def __init__(self) -> None:
pass
def skip_data_noises(self, result_dict):
"""
This function skips the data noises, including overlap blocks, header, footer, watermark, vertical margin note, title
"""
filtered_result_dict = {}
for page_id, blocks in result_dict.items():
if page_id.startswith("page_"):
filtered_blocks = {}
for block_id, block in blocks.items():
if block_id.startswith("block_"):
if any(
block.get(key, 0)
for key in [
"is_overlap",
"is_header",
"is_footer",
"is_watermark",
"is_vertical_margin_note",
"is_block_title",
]
):
continue
filtered_blocks[block_id] = block
if filtered_blocks:
filtered_result_dict[page_id] = filtered_blocks
return filtered_result_dict
magic_pdf/para/draw.py deleted 100644 → 0
View file @ 4df1eb74
from magic_pdf.libs.commons import fitz
from magic_pdf.para.commons import *
if sys.version_info[0] >= 3:
sys.stdout.reconfigure(encoding="utf-8") # type: ignore
class DrawAnnos:
"""
This class draws annotations on the pdf file
----------------------------------------
Color Code
----------------------------------------
Red: (1, 0, 0)
Green: (0, 1, 0)
Blue: (0, 0, 1)
Yellow: (1, 1, 0) - mix of red and green
Cyan: (0, 1, 1) - mix of green and blue
Magenta: (1, 0, 1) - mix of red and blue
White: (1, 1, 1) - red, green and blue full intensity
Black: (0, 0, 0) - no color component whatsoever
Gray: (0.5, 0.5, 0.5) - equal and medium intensity of red, green and blue color components
Orange: (1, 0.65, 0) - maximum intensity of red, medium intensity of green, no blue component
"""
def __init__(self) -> None:
pass
def __is_nested_list(self, lst):
"""
This function returns True if the given list is a nested list of any degree.
"""
if isinstance(lst, list):
return any(self.__is_nested_list(i) for i in lst) or any(isinstance(i, list) for i in lst)
return False
def __valid_rect(self, bbox):
# Ensure that the rectangle is not empty or invalid
if isinstance(bbox[0], list):
return False # It's a nested list, hence it can't be valid rect
else:
return bbox[0] < bbox[2] and bbox[1] < bbox[3]
def __draw_nested_boxes(self, page, nested_bbox, color=(0, 1, 1)):
"""
This function draws the nested boxes
Parameters
----------
page : fitz.Page
page
nested_bbox : list
nested bbox
color : tuple
color, by default (0, 1, 1) # draw with cyan color for combined paragraph
"""
if self.__is_nested_list(nested_bbox): # If it's a nested list
for bbox in nested_bbox:
self.__draw_nested_boxes(page, bbox, color) # Recursively call the function
elif self.__valid_rect(nested_bbox): # If valid rectangle
para_rect = fitz.Rect(nested_bbox)
para_anno = page.add_rect_annot(para_rect)
para_anno.set_colors(stroke=color) # draw with cyan color for combined paragraph
para_anno.set_border(width=1)
para_anno.update()
def draw_annos(self, input_pdf_path, pdf_dic, output_pdf_path):
pdf_doc = open_pdf(input_pdf_path)
if pdf_dic is None:
pdf_dic = {}
if output_pdf_path is None:
output_pdf_path = input_pdf_path.replace(".pdf", "_anno.pdf")
for page_id, page in enumerate(pdf_doc): # type: ignore
page_key = f"page_{page_id}"
for ele_key, ele_data in pdf_dic[page_key].items():
if ele_key == "para_blocks":
para_blocks = ele_data
for para_block in para_blocks:
if "paras" in para_block.keys():
paras = para_block["paras"]
for para_key, para_content in paras.items():
para_bbox = para_content["para_bbox"]
# print(f"para_bbox: {para_bbox}")
# print(f"is a nested list: {self.__is_nested_list(para_bbox)}")
if self.__is_nested_list(para_bbox) and len(para_bbox) > 1:
color = (0, 1, 1)
self.__draw_nested_boxes(
page, para_bbox, color
) # draw with cyan color for combined paragraph
else:
if self.__valid_rect(para_bbox):
para_rect = fitz.Rect(para_bbox)
para_anno = page.add_rect_annot(para_rect)
para_anno.set_colors(stroke=(0, 1, 0)) # draw with green color for normal paragraph
para_anno.set_border(width=0.5)
para_anno.update()
is_para_title = para_content["is_para_title"]
if is_para_title:
if self.__is_nested_list(para_content["para_bbox"]) and len(para_content["para_bbox"]) > 1:
color = (0, 0, 1)
self.__draw_nested_boxes(
page, para_content["para_bbox"], color
) # draw with cyan color for combined title
else:
if self.__valid_rect(para_content["para_bbox"]):
para_rect = fitz.Rect(para_content["para_bbox"])
if self.__valid_rect(para_content["para_bbox"]):
para_anno = page.add_rect_annot(para_rect)
para_anno.set_colors(stroke=(0, 0, 1)) # draw with blue color for normal title
para_anno.set_border(width=0.5)
para_anno.update()
pdf_doc.save(output_pdf_path)
pdf_doc.close()
magic_pdf/para/exceptions.py deleted 100644 → 0
View file @ 4df1eb74
class DenseSingleLineBlockException(Exception):
"""
This class defines the exception type for dense single line-block.
"""
def __init__(self, message="DenseSingleLineBlockException"):
self.message = message
super().__init__(self.message)
def __str__(self):
return f"{self.message}"
def __repr__(self):
return f"{self.message}"
class TitleDetectionException(Exception):
"""
This class defines the exception type for title detection.
"""
def __init__(self, message="TitleDetectionException"):
self.message = message
super().__init__(self.message)
def __str__(self):
return f"{self.message}"
def __repr__(self):
return f"{self.message}"
class TitleLevelException(Exception):
"""
This class defines the exception type for title level.
"""
def __init__(self, message="TitleLevelException"):
self.message = message
super().__init__(self.message)
def __str__(self):
return f"{self.message}"
def __repr__(self):
return f"{self.message}"
class ParaSplitException(Exception):
"""
This class defines the exception type for paragraph splitting.
"""
def __init__(self, message="ParaSplitException"):
self.message = message
super().__init__(self.message)
def __str__(self):
return f"{self.message}"
def __repr__(self):
return f"{self.message}"
class ParaMergeException(Exception):
"""
This class defines the exception type for paragraph merging.
"""
def __init__(self, message="ParaMergeException"):
self.message = message
super().__init__(self.message)
def __str__(self):
return f"{self.message}"
def __repr__(self):
return f"{self.message}"
class DiscardByException:
"""
This class discards pdf files by exception
"""
def __init__(self) -> None:
pass
def discard_by_single_line_block(self, pdf_dic, exception: DenseSingleLineBlockException):
"""
This function discards pdf files by single line block exception
Parameters
----------
pdf_dic : dict
pdf dictionary
exception : str
exception message
Returns
-------
error_message : str
"""
exception_page_nums = 0
page_num = 0
for page_id, page in pdf_dic.items():
if page_id.startswith("page_"):
page_num += 1
if "preproc_blocks" in page.keys():
preproc_blocks = page["preproc_blocks"]
all_single_line_blocks = []
for block in preproc_blocks:
if len(block["lines"]) == 1:
all_single_line_blocks.append(block)
if len(preproc_blocks) > 0 and len(all_single_line_blocks) / len(preproc_blocks) > 0.9:
exception_page_nums += 1
if page_num == 0:
return None
if exception_page_nums / page_num > 0.1: # Low ratio means basically, whenever this is the case, it is discarded
return exception.message
return None
def discard_by_title_detection(self, pdf_dic, exception: TitleDetectionException):
"""
This function discards pdf files by title detection exception
Parameters
----------
pdf_dic : dict
pdf dictionary
exception : str
exception message
Returns
-------
error_message : str
"""
# return exception.message
return None
def discard_by_title_level(self, pdf_dic, exception: TitleLevelException):
"""
This function discards pdf files by title level exception
Parameters
----------
pdf_dic : dict
pdf dictionary
exception : str
exception message
Returns
-------
error_message : str
"""
# return exception.message
return None
def discard_by_split_para(self, pdf_dic, exception: ParaSplitException):
"""
This function discards pdf files by split para exception
Parameters
----------
pdf_dic : dict
pdf dictionary
exception : str
exception message
Returns
-------
error_message : str
"""
# return exception.message
return None
def discard_by_merge_para(self, pdf_dic, exception: ParaMergeException):
"""
This function discards pdf files by merge para exception
Parameters
----------
pdf_dic : dict
pdf dictionary
exception : str
exception message
Returns
-------
error_message : str
"""
# return exception.message
return None
magic_pdf/para/layout_match_processor.py deleted 100644 → 0
View file @ 4df1eb74
import math
from magic_pdf.para.commons import *
if sys.version_info[0] >= 3:
sys.stdout.reconfigure(encoding="utf-8") # type: ignore
class LayoutFilterProcessor:
def __init__(self) -> None:
pass
def batch_process_blocks(self, pdf_dict):
for page_id, blocks in pdf_dict.items():
if page_id.startswith("page_"):
if "layout_bboxes" in blocks.keys() and "para_blocks" in blocks.keys():
layout_bbox_objs = blocks["layout_bboxes"]
if layout_bbox_objs is None:
continue
layout_bboxes = [bbox_obj["layout_bbox"] for bbox_obj in layout_bbox_objs]
# Use math.ceil function to enlarge each value of x0, y0, x1, y1 of each layout_bbox
layout_bboxes = [
[math.ceil(x0), math.ceil(y0), math.ceil(x1), math.ceil(y1)] for x0, y0, x1, y1 in layout_bboxes
]
para_blocks = blocks["para_blocks"]
if para_blocks is None:
continue
for lb_bbox in layout_bboxes:
for i, para_block in enumerate(para_blocks):
para_bbox = para_block["bbox"]
para_blocks[i]["in_layout"] = 0
if is_in_bbox(para_bbox, lb_bbox):
para_blocks[i]["in_layout"] = 1
blocks["para_blocks"] = para_blocks
return pdf_dict
magic_pdf/para/para_split.py deleted 100644 → 0
View file @ 4df1eb74
import numpy as np
from loguru import logger
from sklearn.cluster import DBSCAN
from magic_pdf.config.ocr_content_type import ContentType
from magic_pdf.libs.boxbase import \
_is_in_or_part_overlap_with_area_ratio as is_in_layout
LINE_STOP_FLAG = ['.', '!', '?', '。', '！', '？', '：', ':', ')', '）', ';']
INLINE_EQUATION = ContentType.InlineEquation
INTERLINE_EQUATION = ContentType.InterlineEquation
TEXT = ContentType.Text
def __get_span_text(span):
c = span.get('content', '')
if len(c) == 0:
c = span.get('image_path', '')
return c
def __detect_list_lines(lines, new_layout_bboxes, lang):
"""探测是否包含了列表，并且把列表的行分开.
这样的段落特点是，顶格字母大写/数字，紧跟着几行缩进的。缩进的行首字母含小写的。
"""
def find_repeating_patterns(lst):
indices = []
ones_indices = []
i = 0
while i < len(lst) - 1: # 确保余下元素至少有2个
if lst[i] == 1 and lst[i + 1] in [2, 3]: # 额外检查以防止连续出现的1
start = i
ones_in_this_interval = [i]
i += 1
while i < len(lst) and lst[i] in [2, 3]:
i += 1
# 验证下一个序列是否符合条件
if (
i < len(lst) - 1
and lst[i] == 1
and lst[i + 1] in [2, 3]
and lst[i - 1] in [2, 3]
):
while i < len(lst) and lst[i] in [1, 2, 3]:
if lst[i] == 1:
ones_in_this_interval.append(i)
i += 1
indices.append((start, i - 1))
ones_indices.append(ones_in_this_interval)
else:
i += 1
else:
i += 1
return indices, ones_indices
"""===================="""
def split_indices(slen, index_array):
result = []
last_end = 0
for start, end in sorted(index_array):
if start > last_end:
# 前一个区间结束到下一个区间开始之间的部分标记为"text"
result.append(('text', last_end, start - 1))
# 区间内标记为"list"
result.append(('list', start, end))
last_end = end + 1
if last_end < slen:
# 如果最后一个区间结束后还有剩余的字符串，将其标记为"text"
result.append(('text', last_end, slen - 1))
return result
"""===================="""
if lang != 'en':
return lines, None
else:
total_lines = len(lines)
line_fea_encode = []
"""
对每一行进行特征编码，编码规则如下：
1. 如果行顶格，且大写字母开头或者数字开头，编码为1
2. 如果顶格，其他非大写开头编码为4
3. 如果非顶格，首字符大写，编码为2
4. 如果非顶格，首字符非大写编码为3
"""
for l in lines: # noqa: E741
first_char = __get_span_text(l['spans'][0])[0]
layout_left = __find_layout_bbox_by_line(l['bbox'], new_layout_bboxes)[0]
if l['bbox'][0] == layout_left:
if first_char.isupper() or first_char.isdigit():
line_fea_encode.append(1)
else:
line_fea_encode.append(4)
else:
if first_char.isupper():
line_fea_encode.append(2)
else:
line_fea_encode.append(3)
# 然后根据编码进行分段, 选出来 1,2,3连续出现至少2次的行，认为是列表。
list_indice, list_start_idx = find_repeating_patterns(line_fea_encode)
if len(list_indice) > 0:
logger.info(f'发现了列表，列表行数：{list_indice}， {list_start_idx}')
# TODO check一下这个特列表里缩进的行左侧是不是对齐的。
for start, end in list_indice:
for i in range(start, end + 1):
if i > 0:
if line_fea_encode[i] == 4:
logger.info(f'列表行的第{i}行不是顶格的')
break
else:
logger.info(f'列表行的第{start}到第{end}行是列表')
return split_indices(total_lines, list_indice), list_start_idx
def __valign_lines(blocks, layout_bboxes):
"""在一个layoutbox内对齐行的左侧和右侧。 扫描行的左侧和右侧，如果x0,
x1差距不超过一个阈值，就强行对齐到所处layout的左右两侧（和layout有一段距离）。
3是个经验值，TODO，计算得来，可以设置为1.5个正文字符。"""
min_distance = 3
min_sample = 2
new_layout_bboxes = []
for layout_box in layout_bboxes:
blocks_in_layoutbox = [
b for b in blocks if is_in_layout(b['bbox'], layout_box['layout_bbox'])
]
if len(blocks_in_layoutbox) == 0:
continue
x0_lst = np.array(
[
[line['bbox'][0], 0]
for block in blocks_in_layoutbox
for line in block['lines']
]
)
x1_lst = np.array(
[
[line['bbox'][2], 0]
for block in blocks_in_layoutbox
for line in block['lines']
]
)
x0_clusters = DBSCAN(eps=min_distance, min_samples=min_sample).fit(x0_lst)
x1_clusters = DBSCAN(eps=min_distance, min_samples=min_sample).fit(x1_lst)
x0_uniq_label = np.unique(x0_clusters.labels_)
x1_uniq_label = np.unique(x1_clusters.labels_)
x0_2_new_val = {} # 存储旧值对应的新值映射
x1_2_new_val = {}
for label in x0_uniq_label:
if label == -1:
continue
x0_index_of_label = np.where(x0_clusters.labels_ == label)
x0_raw_val = x0_lst[x0_index_of_label][:, 0]
x0_new_val = np.min(x0_lst[x0_index_of_label][:, 0])
x0_2_new_val.update({idx: x0_new_val for idx in x0_raw_val})
for label in x1_uniq_label:
if label == -1:
continue
x1_index_of_label = np.where(x1_clusters.labels_ == label)
x1_raw_val = x1_lst[x1_index_of_label][:, 0]
x1_new_val = np.max(x1_lst[x1_index_of_label][:, 0])
x1_2_new_val.update({idx: x1_new_val for idx in x1_raw_val})
for block in blocks_in_layoutbox:
for line in block['lines']:
x0, x1 = line['bbox'][0], line['bbox'][2]
if x0 in x0_2_new_val:
line['bbox'][0] = int(x0_2_new_val[x0])
if x1 in x1_2_new_val:
line['bbox'][2] = int(x1_2_new_val[x1])
# 其余对不齐的保持不动
# 由于修改了block里的line长度，现在需要重新计算block的bbox
for block in blocks_in_layoutbox:
block['bbox'] = [
min([line['bbox'][0] for line in block['lines']]),
min([line['bbox'][1] for line in block['lines']]),
max([line['bbox'][2] for line in block['lines']]),
max([line['bbox'][3] for line in block['lines']]),
]
"""新计算layout的bbox，因为block的bbox变了。"""
layout_x0 = min([block['bbox'][0] for block in blocks_in_layoutbox])
layout_y0 = min([block['bbox'][1] for block in blocks_in_layoutbox])
layout_x1 = max([block['bbox'][2] for block in blocks_in_layoutbox])
layout_y1 = max([block['bbox'][3] for block in blocks_in_layoutbox])
new_layout_bboxes.append([layout_x0, layout_y0, layout_x1, layout_y1])
return new_layout_bboxes
def __align_text_in_layout(blocks, layout_bboxes):
"""由于ocr出来的line，有时候会在前后有一段空白，这个时候需要对文本进行对齐，超出的部分被layout左右侧截断。"""
for layout in layout_bboxes:
lb = layout['layout_bbox']
blocks_in_layoutbox = [b for b in blocks if is_in_layout(b['bbox'], lb)]
if len(blocks_in_layoutbox) == 0:
continue
for block in blocks_in_layoutbox:
for line in block['lines']:
x0, x1 = line['bbox'][0], line['bbox'][2]
if x0 < lb[0]:
line['bbox'][0] = lb[0]
if x1 > lb[2]:
line['bbox'][2] = lb[2]
def __common_pre_proc(blocks, layout_bboxes):
"""不分语言的，对文本进行预处理."""
# __add_line_period(blocks, layout_bboxes)
__align_text_in_layout(blocks, layout_bboxes)
aligned_layout_bboxes = __valign_lines(blocks, layout_bboxes)
return aligned_layout_bboxes
def __pre_proc_zh_blocks(blocks, layout_bboxes):
"""对中文文本进行分段预处理."""
pass
def __pre_proc_en_blocks(blocks, layout_bboxes):
"""对英文文本进行分段预处理."""
pass
def __group_line_by_layout(blocks, layout_bboxes, lang='en'):
"""每个layout内的行进行聚合."""
# 因为只是一个block一行目前, 一个block就是一个段落
lines_group = []
for lyout in layout_bboxes:
lines = [
line
for block in blocks
if is_in_layout(block['bbox'], lyout['layout_bbox'])
for line in block['lines']
]
lines_group.append(lines)
return lines_group
def __split_para_in_layoutbox(lines_group, new_layout_bbox, lang='en', char_avg_len=10):
"""
lines_group 进行行分段——layout内部进行分段。lines_group内每个元素是一个Layoutbox内的所有行。
1. 先计算每个group的左右边界。
2. 然后根据行末尾特征进行分段。
末尾特征：以句号等结束符结尾。并且距离右侧边界有一定距离。
且下一行开头不留空白。
"""
list_info = [] # 这个layout最后是不是列表,记录每一个layout里是不是列表开头，列表结尾
layout_paras = []
right_tail_distance = 1.5 * char_avg_len
for lines in lines_group:
paras = []
total_lines = len(lines)
if total_lines == 0:
continue # 0行无需处理
if total_lines == 1: # 1行无法分段。
layout_paras.append([lines])
list_info.append([False, False])
continue
"""在进入到真正的分段之前，要对文字块从统计维度进行对齐方式的探测，
对齐方式分为以下：
1. 左对齐的文本块(特点是左侧顶格，或者左侧不顶格但是右侧顶格的行数大于非顶格的行数，顶格的首字母有大写也有小写)
1) 右侧对齐的行，单独成一段
2) 中间对齐的行，按照字体/行高聚合成一段
2. 左对齐的列表块（其特点是左侧顶格的行数小于等于非顶格的行数，非定格首字母会有小写，顶格90%是大写。并且左侧顶格行数大于1，大于1是为了这种模式连续出现才能称之为列表）
这样的文本块，顶格的为一个段落开头，紧随其后非顶格的行属于这个段落。
"""
text_segments, list_start_line = __detect_list_lines(
lines, new_layout_bbox, lang
)
"""根据list_range，把lines分成几个部分
"""
layout_right = __find_layout_bbox_by_line(lines[0]['bbox'], new_layout_bbox)[2]
layout_left = __find_layout_bbox_by_line(lines[0]['bbox'], new_layout_bbox)[0]
para = [] # 元素是line
layout_list_info = [
False,
False,
] # 这个layout最后是不是列表,记录每一个layout里是不是列表开头，列表结尾
for content_type, start, end in text_segments:
if content_type == 'list':
for i, line in enumerate(lines[start : end + 1]):
line_x0 = line['bbox'][0]
if line_x0 == layout_left: # 列表开头
if len(para) > 0:
paras.append(para)
para = []
para.append(line)
else:
para.append(line)
if len(para) > 0:
paras.append(para)
para = []
if start == 0:
layout_list_info[0] = True
if end == total_lines - 1:
layout_list_info[1] = True
else: # 是普通文本
for i, line in enumerate(lines[start : end + 1]):
# 如果i有下一行，那么就要根据下一行位置综合判断是否要分段。如果i之后没有行，那么只需要判断i行自己的结尾特征。
cur_line_type = line['spans'][-1]['type']
next_line = lines[i + 1] if i < total_lines - 1 else None
if cur_line_type in [TEXT, INLINE_EQUATION]:
if line['bbox'][2] < layout_right - right_tail_distance:
para.append(line)
paras.append(para)
para = []
elif (
line['bbox'][2] >= layout_right - right_tail_distance
and next_line
and next_line['bbox'][0] == layout_left
): # 现在这行到了行尾沾满，下一行存在且顶格。
para.append(line)
else:
para.append(line)
paras.append(para)
para = []
else: # 其他，图片、表格、行间公式，各自占一段
if len(para) > 0: # 先把之前的段落加入到结果中
paras.append(para)
para = []
paras.append(
[line]
) # 再把当前行加入到结果中。当前行为行间公式、图、表等。
para = []
if len(para) > 0:
paras.append(para)
para = []
list_info.append(layout_list_info)
layout_paras.append(paras)
paras = []
return layout_paras, list_info
def __connect_list_inter_layout(
layout_paras, new_layout_bbox, layout_list_info, page_num, lang
):
"""如果上个layout的最后一个段落是列表，下一个layout的第一个段落也是列表，那么将他们连接起来。 TODO
因为没有区分列表和段落，所以这个方法暂时不实现。
根据layout_list_info判断是不是列表。，下个layout的第一个段如果不是列表，那么看他们是否有几行都有相同的缩进。"""
if (
len(layout_paras) == 0 or len(layout_list_info) == 0
): # 0的时候最后的return 会出错
return layout_paras, [False, False]
for i in range(1, len(layout_paras)):
pre_layout_list_info = layout_list_info[i - 1]
next_layout_list_info = layout_list_info[i]
pre_last_para = layout_paras[i - 1][-1]
next_paras = layout_paras[i]
if (
pre_layout_list_info[1] and not next_layout_list_info[0]
): # 前一个是列表结尾，后一个是非列表开头，此时检测是否有相同的缩进
logger.info(f'连接page {page_num} 内的list')
# 向layout_paras[i] 寻找开头具有相同缩进的连续的行
may_list_lines = []
for j in range(len(next_paras)):
line = next_paras[j]
if len(line) == 1: # 只可能是一行，多行情况再需要分析了
if (
line[0]['bbox'][0]
> __find_layout_bbox_by_line(line[0]['bbox'], new_layout_bbox)[
0
]
):
may_list_lines.append(line[0])
else:
break
else:
break
# 如果这些行的缩进是相等的，那么连到上一个layout的最后一个段落上。
if (
len(may_list_lines) > 0
and len(set([x['bbox'][0] for x in may_list_lines])) == 1
):
pre_last_para.extend(may_list_lines)
layout_paras[i] = layout_paras[i][len(may_list_lines) :]
return layout_paras, [
layout_list_info[0][0],
layout_list_info[-1][1],
] # 同时还返回了这个页面级别的开头、结尾是不是列表的信息
def __connect_list_inter_page(
pre_page_paras,
next_page_paras,
pre_page_layout_bbox,
next_page_layout_bbox,
pre_page_list_info,
next_page_list_info,
page_num,
lang,
):
"""如果上个layout的最后一个段落是列表，下一个layout的第一个段落也是列表，那么将他们连接起来。 TODO
因为没有区分列表和段落，所以这个方法暂时不实现。
根据layout_list_info判断是不是列表。，下个layout的第一个段如果不是列表，那么看他们是否有几行都有相同的缩进。"""
if (
len(pre_page_paras) == 0 or len(next_page_paras) == 0
): # 0的时候最后的return 会出错
return False
if (
pre_page_list_info[1] and not next_page_list_info[0]
): # 前一个是列表结尾，后一个是非列表开头，此时检测是否有相同的缩进
logger.info(f'连接page {page_num} 内的list')
# 向layout_paras[i] 寻找开头具有相同缩进的连续的行
may_list_lines = []
for j in range(len(next_page_paras[0])):
line = next_page_paras[0][j]
if len(line) == 1: # 只可能是一行，多行情况再需要分析了
if (
line[0]['bbox'][0]
> __find_layout_bbox_by_line(
line[0]['bbox'], next_page_layout_bbox
)[0]
):
may_list_lines.append(line[0])
else:
break
else:
break
# 如果这些行的缩进是相等的，那么连到上一个layout的最后一个段落上。
if (
len(may_list_lines) > 0
and len(set([x['bbox'][0] for x in may_list_lines])) == 1
):
pre_page_paras[-1].append(may_list_lines)
next_page_paras[0] = next_page_paras[0][len(may_list_lines) :]
return True
return False
def __find_layout_bbox_by_line(line_bbox, layout_bboxes):
"""根据line找到所在的layout."""
for layout in layout_bboxes:
if is_in_layout(line_bbox, layout):
return layout
return None
def __connect_para_inter_layoutbox(layout_paras, new_layout_bbox, lang):
"""
layout之间进行分段。
主要是计算前一个layOut的最后一行和后一个layout的第一行是否可以连接。
连接的条件需要同时满足：
1. 上一个layout的最后一行沾满整个行。并且没有结尾符号。
2. 下一行开头不留空白。
"""
connected_layout_paras = []
if len(layout_paras) == 0:
return connected_layout_paras
connected_layout_paras.append(layout_paras[0])
for i in range(1, len(layout_paras)):
try:
if (
len(layout_paras[i]) == 0 or len(layout_paras[i - 1]) == 0
): # TODO 考虑连接问题，
continue
pre_last_line = layout_paras[i - 1][-1][-1]
next_first_line = layout_paras[i][0][0]
except Exception:
logger.error(f'page layout {i} has no line')
continue
pre_last_line_text = ''.join(
[__get_span_text(span) for span in pre_last_line['spans']]
)
pre_last_line_type = pre_last_line['spans'][-1]['type']
next_first_line_text = ''.join(
[__get_span_text(span) for span in next_first_line['spans']]
)
next_first_line_type = next_first_line['spans'][0]['type']
if pre_last_line_type not in [
TEXT,
INLINE_EQUATION,
] or next_first_line_type not in [TEXT, INLINE_EQUATION]:
connected_layout_paras.append(layout_paras[i])
continue
pre_x2_max = __find_layout_bbox_by_line(pre_last_line['bbox'], new_layout_bbox)[
2
]
next_x0_min = __find_layout_bbox_by_line(
next_first_line['bbox'], new_layout_bbox
)[0]
pre_last_line_text = pre_last_line_text.strip()
next_first_line_text = next_first_line_text.strip()
if (
pre_last_line['bbox'][2] == pre_x2_max
and pre_last_line_text[-1] not in LINE_STOP_FLAG
and next_first_line['bbox'][0] == next_x0_min
): # 前面一行沾满了整个行，并且没有结尾符号.下一行没有空白开头。
"""连接段落条件成立，将前一个layout的段落和后一个layout的段落连接。"""
connected_layout_paras[-1][-1].extend(layout_paras[i][0])
layout_paras[i].pop(
0
) # 删除后一个layout的第一个段落， 因为他已经被合并到前一个layout的最后一个段落了。
if len(layout_paras[i]) == 0:
layout_paras.pop(i)
else:
connected_layout_paras.append(layout_paras[i])
else:
"""连接段落条件不成立，将前一个layout的段落加入到结果中。"""
connected_layout_paras.append(layout_paras[i])
return connected_layout_paras
def __connect_para_inter_page(
pre_page_paras,
next_page_paras,
pre_page_layout_bbox,
next_page_layout_bbox,
page_num,
lang,
):
"""
连接起来相邻两个页面的段落——前一个页面最后一个段落和后一个页面的第一个段落。
是否可以连接的条件：
1. 前一个页面的最后一个段落最后一行沾满整个行。并且没有结尾符号。
2. 后一个页面的第一个段落第一行没有空白开头。
"""
# 有的页面可能压根没有文字
if (
len(pre_page_paras) == 0
or len(next_page_paras) == 0
or len(pre_page_paras[0]) == 0
or len(next_page_paras[0]) == 0
): # TODO [[]]为什么出现在pre_page_paras里？
return False
pre_last_para = pre_page_paras[-1][-1]
next_first_para = next_page_paras[0][0]
pre_last_line = pre_last_para[-1]
next_first_line = next_first_para[0]
pre_last_line_text = ''.join(
[__get_span_text(span) for span in pre_last_line['spans']]
)
pre_last_line_type = pre_last_line['spans'][-1]['type']
next_first_line_text = ''.join(
[__get_span_text(span) for span in next_first_line['spans']]
)
next_first_line_type = next_first_line['spans'][0]['type']
if pre_last_line_type not in [
TEXT,
INLINE_EQUATION,
] or next_first_line_type not in [
TEXT,
INLINE_EQUATION,
]: # TODO，真的要做好，要考虑跨table, image, 行间的情况
# 不是文本，不连接
return False
pre_x2_max = __find_layout_bbox_by_line(
pre_last_line['bbox'], pre_page_layout_bbox
)[2]
next_x0_min = __find_layout_bbox_by_line(
next_first_line['bbox'], next_page_layout_bbox
)[0]
pre_last_line_text = pre_last_line_text.strip()
next_first_line_text = next_first_line_text.strip()
if (
pre_last_line['bbox'][2] == pre_x2_max
and pre_last_line_text[-1] not in LINE_STOP_FLAG
and next_first_line['bbox'][0] == next_x0_min
): # 前面一行沾满了整个行，并且没有结尾符号.下一行没有空白开头。
"""连接段落条件成立，将前一个layout的段落和后一个layout的段落连接。"""
pre_last_para.extend(next_first_para)
next_page_paras[0].pop(
0
) # 删除后一个页面的第一个段落， 因为他已经被合并到前一个页面的最后一个段落了。
return True
else:
return False
def find_consecutive_true_regions(input_array):
start_index = None # 连续True区域的起始索引
regions = [] # 用于保存所有连续True区域的起始和结束索引
for i in range(len(input_array)):
# 如果我们找到了一个True值，并且当前并没有在连续True区域中
if input_array[i] and start_index is None:
start_index = i # 记录连续True区域的起始索引
# 如果我们找到了一个False值，并且当前在连续True区域中
elif not input_array[i] and start_index is not None:
# 如果连续True区域长度大于1，那么将其添加到结果列表中
if i - start_index > 1:
regions.append((start_index, i - 1))
start_index = None # 重置起始索引
# 如果最后一个元素是True，那么需要将最后一个连续True区域加入到结果列表中
if start_index is not None and len(input_array) - start_index > 1:
regions.append((start_index, len(input_array) - 1))
return regions
def __connect_middle_align_text(
page_paras, new_layout_bbox, page_num, lang, debug_mode
):
"""
找出来中间对齐的连续单行文本，如果连续行高度相同，那么合并为一个段落。
一个line居中的条件是：
1. 水平中心点跨越layout的中心点。
2. 左右两侧都有空白
"""
for layout_i, layout_para in enumerate(page_paras):
layout_box = new_layout_bbox[layout_i]
single_line_paras_tag = []
for i in range(len(layout_para)):
single_line_paras_tag.append(
len(layout_para[i]) == 1
and layout_para[i][0]['spans'][0]['type'] == TEXT
)
"""找出来连续的单行文本，如果连续行高度相同，那么合并为一个段落。"""
consecutive_single_line_indices = find_consecutive_true_regions(
single_line_paras_tag
)
if len(consecutive_single_line_indices) > 0:
index_offset = 0
"""检查这些行是否是高度相同的，居中的"""
for start, end in consecutive_single_line_indices:
start += index_offset
end += index_offset
line_hi = np.array(
[
line[0]['bbox'][3] - line[0]['bbox'][1]
for line in layout_para[start : end + 1]
]
)
first_line_text = ''.join(
[__get_span_text(span) for span in layout_para[start][0]['spans']]
)
if 'Table' in first_line_text or 'Figure' in first_line_text:
pass
if debug_mode:
logger.debug(line_hi.std())
if line_hi.std() < 2:
"""行高度相同，那么判断是否居中."""
all_left_x0 = [
line[0]['bbox'][0] for line in layout_para[start : end + 1]
]
all_right_x1 = [
line[0]['bbox'][2] for line in layout_para[start : end + 1]
]
layout_center = (layout_box[0] + layout_box[2]) / 2
if (
all(
[
x0 < layout_center < x1
for x0, x1 in zip(all_left_x0, all_right_x1)
]
)
and not all([x0 == layout_box[0] for x0 in all_left_x0])
and not all([x1 == layout_box[2] for x1 in all_right_x1])
):
merge_para = [l[0] for l in layout_para[start : end + 1]] # noqa: E741
para_text = ''.join(
[
__get_span_text(span)
for line in merge_para
for span in line['spans']
]
)
if debug_mode:
logger.debug(para_text)
layout_para[start : end + 1] = [merge_para]
index_offset -= end - start
return
def __merge_signle_list_text(page_paras, new_layout_bbox, page_num, lang):
"""找出来连续的单行文本，如果首行顶格，接下来的几个单行段落缩进对齐，那么合并为一个段落。"""
pass
def __do_split_page(blocks, layout_bboxes, new_layout_bbox, page_num, lang):
"""根据line和layout情况进行分段 先实现一个根据行末尾特征分段的简单方法。"""
"""
算法思路：
1. 扫描layout里每一行，找出来行尾距离layout有边界有一定距离的行。
2. 从上述行中找到末尾是句号等可作为断行标志的行。
3. 参照上述行尾特征进行分段。
4. 图、表，目前独占一行，不考虑分段。
"""
if page_num == 343:
pass
lines_group = __group_line_by_layout(blocks, layout_bboxes, lang) # block内分段
layout_paras, layout_list_info = __split_para_in_layoutbox(
lines_group, new_layout_bbox, lang
) # layout内分段
layout_paras2, page_list_info = __connect_list_inter_layout(
layout_paras, new_layout_bbox, layout_list_info, page_num, lang
) # layout之间连接列表段落
connected_layout_paras = __connect_para_inter_layoutbox(
layout_paras2, new_layout_bbox, lang
) # layout间链接段落
return connected_layout_paras, page_list_info
def para_split(pdf_info_dict, debug_mode, lang='en'):
"""根据line和layout情况进行分段."""
new_layout_of_pages = [] # 数组的数组，每个元素是一个页面的layoutS
all_page_list_info = [] # 保存每个页面开头和结尾是否是列表
for page_num, page in pdf_info_dict.items():
blocks = page['preproc_blocks']
layout_bboxes = page['layout_bboxes']
new_layout_bbox = __common_pre_proc(blocks, layout_bboxes)
new_layout_of_pages.append(new_layout_bbox)
splited_blocks, page_list_info = __do_split_page(
blocks, layout_bboxes, new_layout_bbox, page_num, lang
)
all_page_list_info.append(page_list_info)
page['para_blocks'] = splited_blocks
"""连接页面与页面之间的可能合并的段落"""
pdf_infos = list(pdf_info_dict.values())
for page_num, page in enumerate(pdf_info_dict.values()):
if page_num == 0:
continue
pre_page_paras = pdf_infos[page_num - 1]['para_blocks']
next_page_paras = pdf_infos[page_num]['para_blocks']
pre_page_layout_bbox = new_layout_of_pages[page_num - 1]
next_page_layout_bbox = new_layout_of_pages[page_num]
is_conn = __connect_para_inter_page(
pre_page_paras,
next_page_paras,
pre_page_layout_bbox,
next_page_layout_bbox,
page_num,
lang,
)
if debug_mode:
if is_conn:
logger.info(f'连接了第{page_num-1}页和第{page_num}页的段落')
is_list_conn = __connect_list_inter_page(
pre_page_paras,
next_page_paras,
pre_page_layout_bbox,
next_page_layout_bbox,
all_page_list_info[page_num - 1],
all_page_list_info[page_num],
page_num,
lang,
)
if debug_mode:
if is_list_conn:
logger.info(f'连接了第{page_num-1}页和第{page_num}页的列表段落')
"""接下来可能会漏掉一些特别的一些可以合并的内容，对他们进行段落连接
1. 正文中有时出现一个行顶格，接下来几行缩进的情况。
2. 居中的一些连续单行，如果高度相同，那么可能是一个段落。
"""
for page_num, page in enumerate(pdf_info_dict.values()):
page_paras = page['para_blocks']
new_layout_bbox = new_layout_of_pages[page_num]
__connect_middle_align_text(
page_paras, new_layout_bbox, page_num, lang, debug_mode=debug_mode
)
__merge_signle_list_text(page_paras, new_layout_bbox, page_num, lang)
magic_pdf/para/para_split_v2.py deleted 100644 → 0
View file @ 4df1eb74
import copy
import re
import numpy as np
from loguru import logger
from sklearn.cluster import DBSCAN
from magic_pdf.config.constants import * # noqa: F403
from magic_pdf.config.ocr_content_type import BlockType, ContentType
from magic_pdf.libs.boxbase import \
_is_in_or_part_overlap_with_area_ratio as is_in_layout
LINE_STOP_FLAG = ['.', '!', '?', '。', '！', '？', '：', ':', ')', '）', ';']
INLINE_EQUATION = ContentType.InlineEquation
INTERLINE_EQUATION = ContentType.InterlineEquation
TEXT = ContentType.Text
debug_able = False
def __get_span_text(span):
c = span.get('content', '')
if len(c) == 0:
c = span.get('image_path', '')
return c
def __detect_list_lines(lines, new_layout_bboxes, lang):
global debug_able
"""
探测是否包含了列表，并且把列表的行分开.
这样的段落特点是，顶格字母大写/数字，紧跟着几行缩进的。缩进的行首字母含小写的。
"""
def find_repeating_patterns2(lst):
indices = []
ones_indices = []
i = 0
while i < len(lst): # Loop through the entire list
if (
lst[i] == 1
): # If we encounter a '1', we might be at the start of a pattern
start = i
ones_in_this_interval = [i]
i += 1
# Traverse elements that are 1, 2 or 3, until we encounter something else
while i < len(lst) and lst[i] in [1, 2, 3]:
if lst[i] == 1:
ones_in_this_interval.append(i)
i += 1
if len(ones_in_this_interval) > 1 or (
start < len(lst) - 1
and ones_in_this_interval
and lst[start + 1] in [2, 3]
):
indices.append((start, i - 1))
ones_indices.append(ones_in_this_interval)
else:
i += 1
return indices, ones_indices
def find_repeating_patterns(lst):
indices = []
ones_indices = []
i = 0
while i < len(lst) - 1: # 确保余下元素至少有2个
if lst[i] == 1 and lst[i + 1] in [2, 3]: # 额外检查以防止连续出现的1
start = i
ones_in_this_interval = [i]
i += 1
while i < len(lst) and lst[i] in [2, 3]:
i += 1
# 验证下一个序列是否符合条件
if (
i < len(lst) - 1
and lst[i] == 1
and lst[i + 1] in [2, 3]
and lst[i - 1] in [2, 3]
):
while i < len(lst) and lst[i] in [1, 2, 3]:
if lst[i] == 1:
ones_in_this_interval.append(i)
i += 1
indices.append((start, i - 1))
ones_indices.append(ones_in_this_interval)
else:
i += 1
else:
i += 1
return indices, ones_indices
"""===================="""
def split_indices(slen, index_array):
result = []
last_end = 0
for start, end in sorted(index_array):
if start > last_end:
# 前一个区间结束到下一个区间开始之间的部分标记为"text"
result.append(('text', last_end, start - 1))
# 区间内标记为"list"
result.append(('list', start, end))
last_end = end + 1
if last_end < slen:
# 如果最后一个区间结束后还有剩余的字符串，将其标记为"text"
result.append(('text', last_end, slen - 1))
return result
"""===================="""
if lang != 'en':
return lines, None
total_lines = len(lines)
line_fea_encode = []
"""
对每一行进行特征编码，编码规则如下：
1. 如果行顶格，且大写字母开头或者数字开头，编码为1
2. 如果顶格，其他非大写开头编码为4
3. 如果非顶格，首字符大写，编码为2
4. 如果非顶格，首字符非大写编码为3
"""
if len(lines) > 0:
x_map_tag_dict, min_x_tag = cluster_line_x(lines)
for l in lines: # noqa: E741
span_text = __get_span_text(l['spans'][0])
if not span_text:
line_fea_encode.append(0)
continue
first_char = span_text[0]
layout = __find_layout_bbox_by_line(l['bbox'], new_layout_bboxes)
if not layout:
line_fea_encode.append(0)
else:
#
if x_map_tag_dict[round(l['bbox'][0])] == min_x_tag:
# if first_char.isupper() or first_char.isdigit() or not first_char.isalnum():
if not first_char.isalnum() or if_match_reference_list(span_text):
line_fea_encode.append(1)
else:
line_fea_encode.append(4)
else:
if first_char.isupper():
line_fea_encode.append(2)
else:
line_fea_encode.append(3)
# 然后根据编码进行分段, 选出来 1,2,3连续出现至少2次的行，认为是列表。
list_indice, list_start_idx = find_repeating_patterns2(line_fea_encode)
if len(list_indice) > 0:
if debug_able:
logger.info(f'发现了列表，列表行数：{list_indice}， {list_start_idx}')
# TODO check一下这个特列表里缩进的行左侧是不是对齐的。
for start, end in list_indice:
for i in range(start, end + 1):
if i > 0:
if line_fea_encode[i] == 4:
if debug_able:
logger.info(f'列表行的第{i}行不是顶格的')
break
else:
if debug_able:
logger.info(f'列表行的第{start}到第{end}行是列表')
return split_indices(total_lines, list_indice), list_start_idx
def cluster_line_x(lines: list) -> dict:
"""对一个block内所有lines的bbox的x0聚类."""
min_distance = 5
min_sample = 1
x0_lst = np.array([[round(line['bbox'][0]), 0] for line in lines])
x0_clusters = DBSCAN(eps=min_distance, min_samples=min_sample).fit(x0_lst)
x0_uniq_label = np.unique(x0_clusters.labels_)
# x1_lst = np.array([[line['bbox'][2], 0] for line in lines])
x0_2_new_val = {} # 存储旧值对应的新值映射
min_x0 = round(lines[0]['bbox'][0])
for label in x0_uniq_label:
if label == -1:
continue
x0_index_of_label = np.where(x0_clusters.labels_ == label)
x0_raw_val = x0_lst[x0_index_of_label][:, 0]
x0_new_val = np.min(x0_lst[x0_index_of_label][:, 0])
x0_2_new_val.update(
{round(raw_val): round(x0_new_val) for raw_val in x0_raw_val}
)
if x0_new_val < min_x0:
min_x0 = x0_new_val
return x0_2_new_val, min_x0
def if_match_reference_list(text: str) -> bool:
pattern = re.compile(r'^\d+\..*')
if pattern.match(text):
return True
else:
return False
def __valign_lines(blocks, layout_bboxes):
"""在一个layoutbox内对齐行的左侧和右侧。 扫描行的左侧和右侧，如果x0,
x1差距不超过一个阈值，就强行对齐到所处layout的左右两侧（和layout有一段距离）。
3是个经验值，TODO，计算得来，可以设置为1.5个正文字符。"""
min_distance = 3
min_sample = 2
new_layout_bboxes = []
# add bbox_fs for para split calculation
for block in blocks:
block['bbox_fs'] = copy.deepcopy(block['bbox'])
for layout_box in layout_bboxes:
blocks_in_layoutbox = [
b
for b in blocks
if b['type'] == BlockType.Text
and is_in_layout(b['bbox'], layout_box['layout_bbox'])
]
if len(blocks_in_layoutbox) == 0 or len(blocks_in_layoutbox[0]['lines']) == 0:
new_layout_bboxes.append(layout_box['layout_bbox'])
continue
x0_lst = np.array(
[
[line['bbox'][0], 0]
for block in blocks_in_layoutbox
for line in block['lines']
]
)
x1_lst = np.array(
[
[line['bbox'][2], 0]
for block in blocks_in_layoutbox
for line in block['lines']
]
)
x0_clusters = DBSCAN(eps=min_distance, min_samples=min_sample).fit(x0_lst)
x1_clusters = DBSCAN(eps=min_distance, min_samples=min_sample).fit(x1_lst)
x0_uniq_label = np.unique(x0_clusters.labels_)
x1_uniq_label = np.unique(x1_clusters.labels_)
x0_2_new_val = {} # 存储旧值对应的新值映射
x1_2_new_val = {}
for label in x0_uniq_label:
if label == -1:
continue
x0_index_of_label = np.where(x0_clusters.labels_ == label)
x0_raw_val = x0_lst[x0_index_of_label][:, 0]
x0_new_val = np.min(x0_lst[x0_index_of_label][:, 0])
x0_2_new_val.update({idx: x0_new_val for idx in x0_raw_val})
for label in x1_uniq_label:
if label == -1:
continue
x1_index_of_label = np.where(x1_clusters.labels_ == label)
x1_raw_val = x1_lst[x1_index_of_label][:, 0]
x1_new_val = np.max(x1_lst[x1_index_of_label][:, 0])
x1_2_new_val.update({idx: x1_new_val for idx in x1_raw_val})
for block in blocks_in_layoutbox:
for line in block['lines']:
x0, x1 = line['bbox'][0], line['bbox'][2]
if x0 in x0_2_new_val:
line['bbox'][0] = int(x0_2_new_val[x0])
if x1 in x1_2_new_val:
line['bbox'][2] = int(x1_2_new_val[x1])
# 其余对不齐的保持不动
# 由于修改了block里的line长度，现在需要重新计算block的bbox
for block in blocks_in_layoutbox:
if len(block['lines']) > 0:
block['bbox_fs'] = [
min([line['bbox'][0] for line in block['lines']]),
min([line['bbox'][1] for line in block['lines']]),
max([line['bbox'][2] for line in block['lines']]),
max([line['bbox'][3] for line in block['lines']]),
]
"""新计算layout的bbox，因为block的bbox变了。"""
layout_x0 = min([block['bbox_fs'][0] for block in blocks_in_layoutbox])
layout_y0 = min([block['bbox_fs'][1] for block in blocks_in_layoutbox])
layout_x1 = max([block['bbox_fs'][2] for block in blocks_in_layoutbox])
layout_y1 = max([block['bbox_fs'][3] for block in blocks_in_layoutbox])
new_layout_bboxes.append([layout_x0, layout_y0, layout_x1, layout_y1])
return new_layout_bboxes
def __align_text_in_layout(blocks, layout_bboxes):
"""由于ocr出来的line，有时候会在前后有一段空白，这个时候需要对文本进行对齐，超出的部分被layout左右侧截断。"""
for layout in layout_bboxes:
lb = layout['layout_bbox']
blocks_in_layoutbox = [
block
for block in blocks
if block['type'] == BlockType.Text and is_in_layout(block['bbox'], lb)
]
if len(blocks_in_layoutbox) == 0:
continue
for block in blocks_in_layoutbox:
for line in block.get('lines', []):
x0, x1 = line['bbox'][0], line['bbox'][2]
if x0 < lb[0]:
line['bbox'][0] = lb[0]
if x1 > lb[2]:
line['bbox'][2] = lb[2]
def __common_pre_proc(blocks, layout_bboxes):
"""不分语言的，对文本进行预处理."""
# __add_line_period(blocks, layout_bboxes)
__align_text_in_layout(blocks, layout_bboxes)
aligned_layout_bboxes = __valign_lines(blocks, layout_bboxes)
return aligned_layout_bboxes
def __pre_proc_zh_blocks(blocks, layout_bboxes):
"""对中文文本进行分段预处理."""
pass
def __pre_proc_en_blocks(blocks, layout_bboxes):
"""对英文文本进行分段预处理."""
pass
def __group_line_by_layout(blocks, layout_bboxes):
"""每个layout内的行进行聚合."""
# 因为只是一个block一行目前, 一个block就是一个段落
blocks_group = []
for lyout in layout_bboxes:
blocks_in_layout = [
block
for block in blocks
if is_in_layout(block.get('bbox_fs', None), lyout['layout_bbox'])
]
blocks_group.append(blocks_in_layout)
return blocks_group
def __split_para_in_layoutbox(blocks_group, new_layout_bbox, lang='en'):
"""
lines_group 进行行分段——layout内部进行分段。lines_group内每个元素是一个Layoutbox内的所有行。
1. 先计算每个group的左右边界。
2. 然后根据行末尾特征进行分段。
末尾特征：以句号等结束符结尾。并且距离右侧边界有一定距离。
且下一行开头不留空白。
"""
list_info = [] # 这个layout最后是不是列表,记录每一个layout里是不是列表开头，列表结尾
for blocks in blocks_group:
is_start_list = None
is_end_list = None
if len(blocks) == 0:
list_info.append([False, False])
continue
if blocks[0]['type'] != BlockType.Text and blocks[-1]['type'] != BlockType.Text:
list_info.append([False, False])
continue
if blocks[0]['type'] != BlockType.Text:
is_start_list = False
if blocks[-1]['type'] != BlockType.Text:
is_end_list = False
lines = [
line
for block in blocks
if block['type'] == BlockType.Text
for line in block['lines']
]
total_lines = len(lines)
if total_lines == 1 or total_lines == 0:
list_info.append([False, False])
continue
"""在进入到真正的分段之前，要对文字块从统计维度进行对齐方式的探测，
对齐方式分为以下：
1. 左对齐的文本块(特点是左侧顶格，或者左侧不顶格但是右侧顶格的行数大于非顶格的行数，顶格的首字母有大写也有小写)
1) 右侧对齐的行，单独成一段
2) 中间对齐的行，按照字体/行高聚合成一段
2. 左对齐的列表块（其特点是左侧顶格的行数小于等于非顶格的行数，非定格首字母会有小写，顶格90%是大写。并且左侧顶格行数大于1，大于1是为了这种模式连续出现才能称之为列表）
这样的文本块，顶格的为一个段落开头，紧随其后非顶格的行属于这个段落。
"""
text_segments, list_start_line = __detect_list_lines(
lines, new_layout_bbox, lang
)
"""根据list_range，把lines分成几个部分
"""
for list_start in list_start_line:
if len(list_start) > 1:
for i in range(0, len(list_start)):
index = list_start[i] - 1
if index >= 0:
if 'content' in lines[index]['spans'][-1] and lines[index][
'spans'
][-1].get('type', '') not in [
ContentType.InlineEquation,
ContentType.InterlineEquation,
]:
lines[index]['spans'][-1]['content'] += '\n\n'
layout_list_info = [
False,
False,
] # 这个layout最后是不是列表,记录每一个layout里是不是列表开头，列表结尾
for content_type, start, end in text_segments:
if content_type == 'list':
if start == 0 and is_start_list is None:
layout_list_info[0] = True
if end == total_lines - 1 and is_end_list is None:
layout_list_info[1] = True
list_info.append(layout_list_info)
return list_info
def __split_para_lines(lines: list, text_blocks: list) -> list:
text_paras = []
other_paras = []
text_lines = []
for line in lines:
spans_types = [span['type'] for span in line]
if ContentType.Table in spans_types:
other_paras.append([line])
continue
if ContentType.Image in spans_types:
other_paras.append([line])
continue
if ContentType.InterlineEquation in spans_types:
other_paras.append([line])
continue
text_lines.append(line)
for block in text_blocks:
block_bbox = block['bbox']
para = []
for line in text_lines:
bbox = line['bbox']
if is_in_layout(bbox, block_bbox):
para.append(line)
if len(para) > 0:
text_paras.append(para)
paras = other_paras.extend(text_paras)
paras_sorted = sorted(paras, key=lambda x: x[0]['bbox'][1])
return paras_sorted
def __connect_list_inter_layout(
blocks_group, new_layout_bbox, layout_list_info, page_num, lang
):
global debug_able
"""
如果上个layout的最后一个段落是列表，下一个layout的第一个段落也是列表，那么将他们连接起来。 TODO 因为没有区分列表和段落，所以这个方法暂时不实现。
根据layout_list_info判断是不是列表。，下个layout的第一个段如果不是列表，那么看他们是否有几行都有相同的缩进。
"""
if len(blocks_group) == 0 or len(blocks_group) == 0: # 0的时候最后的return 会出错
return blocks_group, [False, False]
for i in range(1, len(blocks_group)):
if len(blocks_group[i]) == 0 or len(blocks_group[i - 1]) == 0:
continue
pre_layout_list_info = layout_list_info[i - 1]
next_layout_list_info = layout_list_info[i]
pre_last_para = blocks_group[i - 1][-1].get('lines', [])
next_paras = blocks_group[i]
next_first_para = next_paras[0]
if (
pre_layout_list_info[1]
and not next_layout_list_info[0]
and next_first_para['type'] == BlockType.Text
): # 前一个是列表结尾，后一个是非列表开头，此时检测是否有相同的缩进
if debug_able:
logger.info(f'连接page {page_num} 内的list')
# 向layout_paras[i] 寻找开头具有相同缩进的连续的行
may_list_lines = []
lines = next_first_para.get('lines', [])
for line in lines:
if (
line['bbox'][0]
> __find_layout_bbox_by_line(line['bbox'], new_layout_bbox)[0]
):
may_list_lines.append(line)
else:
break
# 如果这些行的缩进是相等的，那么连到上一个layout的最后一个段落上。
if (
len(may_list_lines) > 0
and len(set([x['bbox'][0] for x in may_list_lines])) == 1
):
pre_last_para.extend(may_list_lines)
next_first_para['lines'] = next_first_para['lines'][
len(may_list_lines) :
]
return blocks_group, [
layout_list_info[0][0],
layout_list_info[-1][1],
] # 同时还返回了这个页面级别的开头、结尾是不是列表的信息
def __connect_list_inter_page(
pre_page_paras,
next_page_paras,
pre_page_layout_bbox,
next_page_layout_bbox,
pre_page_list_info,
next_page_list_info,
page_num,
lang,
):
"""如果上个layout的最后一个段落是列表，下一个layout的第一个段落也是列表，那么将他们连接起来。 TODO
因为没有区分列表和段落，所以这个方法暂时不实现。
根据layout_list_info判断是不是列表。，下个layout的第一个段如果不是列表，那么看他们是否有几行都有相同的缩进。"""
if (
len(pre_page_paras) == 0 or len(next_page_paras) == 0
): # 0的时候最后的return 会出错
return False
if len(pre_page_paras[-1]) == 0 or len(next_page_paras[0]) == 0:
return False
if (
pre_page_paras[-1][-1]['type'] != BlockType.Text
or next_page_paras[0][0]['type'] != BlockType.Text
):
return False
if (
pre_page_list_info[1] and not next_page_list_info[0]
): # 前一个是列表结尾，后一个是非列表开头，此时检测是否有相同的缩进
if debug_able:
logger.info(f'连接page {page_num} 内的list')
# 向layout_paras[i] 寻找开头具有相同缩进的连续的行
may_list_lines = []
next_page_first_para = next_page_paras[0][0]
if next_page_first_para['type'] == BlockType.Text:
lines = next_page_first_para['lines']
for line in lines:
if (
line['bbox'][0]
> __find_layout_bbox_by_line(line['bbox'], next_page_layout_bbox)[0]
):
may_list_lines.append(line)
else:
break
# 如果这些行的缩进是相等的，那么连到上一个layout的最后一个段落上。
if (
len(may_list_lines) > 0
and len(set([x['bbox'][0] for x in may_list_lines])) == 1
):
# pre_page_paras[-1].append(may_list_lines)
# 下一页合并到上一页最后一段，打一个cross_page的标签
for line in may_list_lines:
for span in line['spans']:
span[CROSS_PAGE] = True # noqa: F405
pre_page_paras[-1][-1]['lines'].extend(may_list_lines)
next_page_first_para['lines'] = next_page_first_para['lines'][
len(may_list_lines) :
]
return True
return False
def __find_layout_bbox_by_line(line_bbox, layout_bboxes):
"""根据line找到所在的layout."""
for layout in layout_bboxes:
if is_in_layout(line_bbox, layout):
return layout
return None
def __connect_para_inter_layoutbox(blocks_group, new_layout_bbox):
"""
layout之间进行分段。
主要是计算前一个layOut的最后一行和后一个layout的第一行是否可以连接。
连接的条件需要同时满足：
1. 上一个layout的最后一行沾满整个行。并且没有结尾符号。
2. 下一行开头不留空白。
"""
connected_layout_blocks = []
if len(blocks_group) == 0:
return connected_layout_blocks
connected_layout_blocks.append(blocks_group[0])
for i in range(1, len(blocks_group)):
try:
if len(blocks_group[i]) == 0:
continue
if len(blocks_group[i - 1]) == 0: # TODO 考虑连接问题，
connected_layout_blocks.append(blocks_group[i])
continue
# text类型的段才需要考虑layout间的合并
if (
blocks_group[i - 1][-1]['type'] != BlockType.Text
or blocks_group[i][0]['type'] != BlockType.Text
):
connected_layout_blocks.append(blocks_group[i])
continue
if (
len(blocks_group[i - 1][-1]['lines']) == 0
or len(blocks_group[i][0]['lines']) == 0
):
connected_layout_blocks.append(blocks_group[i])
continue
pre_last_line = blocks_group[i - 1][-1]['lines'][-1]
next_first_line = blocks_group[i][0]['lines'][0]
except Exception:
logger.error(f'page layout {i} has no line')
continue
pre_last_line_text = ''.join(
[__get_span_text(span) for span in pre_last_line['spans']]
)
pre_last_line_type = pre_last_line['spans'][-1]['type']
next_first_line_text = ''.join(
[__get_span_text(span) for span in next_first_line['spans']]
)
next_first_line_type = next_first_line['spans'][0]['type']
if pre_last_line_type not in [
TEXT,
INLINE_EQUATION,
] or next_first_line_type not in [TEXT, INLINE_EQUATION]:
connected_layout_blocks.append(blocks_group[i])
continue
pre_layout = __find_layout_bbox_by_line(pre_last_line['bbox'], new_layout_bbox)
next_layout = __find_layout_bbox_by_line(
next_first_line['bbox'], new_layout_bbox
)
pre_x2_max = pre_layout[2] if pre_layout else -1
next_x0_min = next_layout[0] if next_layout else -1
pre_last_line_text = pre_last_line_text.strip()
next_first_line_text = next_first_line_text.strip()
if (
pre_last_line['bbox'][2] == pre_x2_max
and pre_last_line_text
and pre_last_line_text[-1] not in LINE_STOP_FLAG
and next_first_line['bbox'][0] == next_x0_min
): # 前面一行沾满了整个行，并且没有结尾符号.下一行没有空白开头。
"""连接段落条件成立，将前一个layout的段落和后一个layout的段落连接。"""
connected_layout_blocks[-1][-1]['lines'].extend(blocks_group[i][0]['lines'])
blocks_group[i][0][
'lines'
] = [] # 删除后一个layout第一个段落中的lines，因为他已经被合并到前一个layout的最后一个段落了
blocks_group[i][0][LINES_DELETED] = True # noqa: F405
# if len(layout_paras[i]) == 0:
# layout_paras.pop(i)
# else:
# connected_layout_paras.append(layout_paras[i])
connected_layout_blocks.append(blocks_group[i])
else:
"""连接段落条件不成立，将前一个layout的段落加入到结果中。"""
connected_layout_blocks.append(blocks_group[i])
return connected_layout_blocks
def __connect_para_inter_page(
pre_page_paras,
next_page_paras,
pre_page_layout_bbox,
next_page_layout_bbox,
page_num,
lang,
):
"""
连接起来相邻两个页面的段落——前一个页面最后一个段落和后一个页面的第一个段落。
是否可以连接的条件：
1. 前一个页面的最后一个段落最后一行沾满整个行。并且没有结尾符号。
2. 后一个页面的第一个段落第一行没有空白开头。
"""
# 有的页面可能压根没有文字
if (
len(pre_page_paras) == 0
or len(next_page_paras) == 0
or len(pre_page_paras[0]) == 0
or len(next_page_paras[0]) == 0
): # TODO [[]]为什么出现在pre_page_paras里？
return False
pre_last_block = pre_page_paras[-1][-1]
next_first_block = next_page_paras[0][0]
if (
pre_last_block['type'] != BlockType.Text
or next_first_block['type'] != BlockType.Text
):
return False
if len(pre_last_block['lines']) == 0 or len(next_first_block['lines']) == 0:
return False
pre_last_para = pre_last_block['lines']
next_first_para = next_first_block['lines']
pre_last_line = pre_last_para[-1]
next_first_line = next_first_para[0]
pre_last_line_text = ''.join(
[__get_span_text(span) for span in pre_last_line['spans']]
)
pre_last_line_type = pre_last_line['spans'][-1]['type']
next_first_line_text = ''.join(
[__get_span_text(span) for span in next_first_line['spans']]
)
next_first_line_type = next_first_line['spans'][0]['type']
if pre_last_line_type not in [
TEXT,
INLINE_EQUATION,
] or next_first_line_type not in [
TEXT,
INLINE_EQUATION,
]: # TODO，真的要做好，要考虑跨table, image, 行间的情况
# 不是文本，不连接
return False
pre_x2_max_bbox = __find_layout_bbox_by_line(
pre_last_line['bbox'], pre_page_layout_bbox
)
if not pre_x2_max_bbox:
return False
next_x0_min_bbox = __find_layout_bbox_by_line(
next_first_line['bbox'], next_page_layout_bbox
)
if not next_x0_min_bbox:
return False
pre_x2_max = pre_x2_max_bbox[2]
next_x0_min = next_x0_min_bbox[0]
pre_last_line_text = pre_last_line_text.strip()
next_first_line_text = next_first_line_text.strip()
if (
pre_last_line['bbox'][2] == pre_x2_max
and pre_last_line_text[-1] not in LINE_STOP_FLAG
and next_first_line['bbox'][0] == next_x0_min
): # 前面一行沾满了整个行，并且没有结尾符号.下一行没有空白开头。
"""连接段落条件成立，将前一个layout的段落和后一个layout的段落连接。"""
# 下一页合并到上一页最后一段，打一个cross_page的标签
for line in next_first_para:
for span in line['spans']:
span[CROSS_PAGE] = True # noqa: F405
pre_last_para.extend(next_first_para)
# next_page_paras[0].pop(0) # 删除后一个页面的第一个段落， 因为他已经被合并到前一个页面的最后一个段落了。
next_page_paras[0][0]['lines'] = []
next_page_paras[0][0][LINES_DELETED] = True # noqa: F405
return True
else:
return False
def find_consecutive_true_regions(input_array):
start_index = None # 连续True区域的起始索引
regions = [] # 用于保存所有连续True区域的起始和结束索引
for i in range(len(input_array)):
# 如果我们找到了一个True值，并且当前并没有在连续True区域中
if input_array[i] and start_index is None:
start_index = i # 记录连续True区域的起始索引
# 如果我们找到了一个False值，并且当前在连续True区域中
elif not input_array[i] and start_index is not None:
# 如果连续True区域长度大于1，那么将其添加到结果列表中
if i - start_index > 1:
regions.append((start_index, i - 1))
start_index = None # 重置起始索引
# 如果最后一个元素是True，那么需要将最后一个连续True区域加入到结果列表中
if start_index is not None and len(input_array) - start_index > 1:
regions.append((start_index, len(input_array) - 1))
return regions
def __connect_middle_align_text(page_paras, new_layout_bbox, page_num, lang):
global debug_able
"""
找出来中间对齐的连续单行文本，如果连续行高度相同，那么合并为一个段落。
一个line居中的条件是：
1. 水平中心点跨越layout的中心点。
2. 左右两侧都有空白
"""
for layout_i, layout_para in enumerate(page_paras):
layout_box = new_layout_bbox[layout_i]
single_line_paras_tag = []
for i in range(len(layout_para)):
# single_line_paras_tag.append(len(layout_para[i]) == 1 and layout_para[i][0]['spans'][0]['type'] == TEXT)
single_line_paras_tag.append(
layout_para[i]['type'] == BlockType.Text
and len(layout_para[i]['lines']) == 1
)
"""找出来连续的单行文本，如果连续行高度相同，那么合并为一个段落。"""
consecutive_single_line_indices = find_consecutive_true_regions(
single_line_paras_tag
)
if len(consecutive_single_line_indices) > 0:
"""检查这些行是否是高度相同的，居中的."""
for start, end in consecutive_single_line_indices:
# start += index_offset
# end += index_offset
line_hi = np.array(
[
block['lines'][0]['bbox'][3] - block['lines'][0]['bbox'][1]
for block in layout_para[start : end + 1]
]
)
first_line_text = ''.join(
[
__get_span_text(span)
for span in layout_para[start]['lines'][0]['spans']
]
)
if 'Table' in first_line_text or 'Figure' in first_line_text:
pass
if debug_able:
logger.info(line_hi.std())
if line_hi.std() < 2:
"""行高度相同，那么判断是否居中."""
all_left_x0 = [
block['lines'][0]['bbox'][0]
for block in layout_para[start : end + 1]
]
all_right_x1 = [
block['lines'][0]['bbox'][2]
for block in layout_para[start : end + 1]
]
layout_center = (layout_box[0] + layout_box[2]) / 2
if (
all(
[
x0 < layout_center < x1
for x0, x1 in zip(all_left_x0, all_right_x1)
]
)
and not all([x0 == layout_box[0] for x0 in all_left_x0])
and not all([x1 == layout_box[2] for x1 in all_right_x1])
):
merge_para = [
block['lines'][0] for block in layout_para[start : end + 1]
]
para_text = ''.join(
[
__get_span_text(span)
for line in merge_para
for span in line['spans']
]
)
if debug_able:
logger.info(para_text)
layout_para[start]['lines'] = merge_para
for i_para in range(start + 1, end + 1):
layout_para[i_para]['lines'] = []
layout_para[i_para][LINES_DELETED] = True # noqa: F405
# layout_para[start:end + 1] = [merge_para]
# index_offset -= end - start
return
def __merge_signle_list_text(page_paras, new_layout_bbox, page_num, lang):
"""找出来连续的单行文本，如果首行顶格，接下来的几个单行段落缩进对齐，那么合并为一个段落。"""
pass
def __do_split_page(blocks, layout_bboxes, new_layout_bbox, page_num, lang):
"""根据line和layout情况进行分段 先实现一个根据行末尾特征分段的简单方法。"""
"""
算法思路：
1. 扫描layout里每一行，找出来行尾距离layout有边界有一定距离的行。
2. 从上述行中找到末尾是句号等可作为断行标志的行。
3. 参照上述行尾特征进行分段。
4. 图、表，目前独占一行，不考虑分段。
"""
blocks_group = __group_line_by_layout(blocks, layout_bboxes) # block内分段
layout_list_info = __split_para_in_layoutbox(
blocks_group, new_layout_bbox, lang
) # layout内分段
blocks_group, page_list_info = __connect_list_inter_layout(
blocks_group, new_layout_bbox, layout_list_info, page_num, lang
) # layout之间连接列表段落
connected_layout_blocks = __connect_para_inter_layoutbox(
blocks_group, new_layout_bbox
) # layout间链接段落
return connected_layout_blocks, page_list_info
def para_split(pdf_info_dict, debug_mode, lang='en'):
global debug_able
debug_able = debug_mode
new_layout_of_pages = [] # 数组的数组，每个元素是一个页面的layoutS
all_page_list_info = [] # 保存每个页面开头和结尾是否是列表
for page_num, page in pdf_info_dict.items():
blocks = copy.deepcopy(page['preproc_blocks'])
layout_bboxes = page['layout_bboxes']
new_layout_bbox = __common_pre_proc(blocks, layout_bboxes)
new_layout_of_pages.append(new_layout_bbox)
splited_blocks, page_list_info = __do_split_page(
blocks, layout_bboxes, new_layout_bbox, page_num, lang
)
all_page_list_info.append(page_list_info)
page['para_blocks'] = splited_blocks
"""连接页面与页面之间的可能合并的段落"""
pdf_infos = list(pdf_info_dict.values())
for page_num, page in enumerate(pdf_info_dict.values()):
if page_num == 0:
continue
pre_page_paras = pdf_infos[page_num - 1]['para_blocks']
next_page_paras = pdf_infos[page_num]['para_blocks']
pre_page_layout_bbox = new_layout_of_pages[page_num - 1]
next_page_layout_bbox = new_layout_of_pages[page_num]
is_conn = __connect_para_inter_page(
pre_page_paras,
next_page_paras,
pre_page_layout_bbox,
next_page_layout_bbox,
page_num,
lang,
)
if debug_able:
if is_conn:
logger.info(f'连接了第{page_num - 1}页和第{page_num}页的段落')
is_list_conn = __connect_list_inter_page(
pre_page_paras,
next_page_paras,
pre_page_layout_bbox,
next_page_layout_bbox,
all_page_list_info[page_num - 1],
all_page_list_info[page_num],
page_num,
lang,
)
if debug_able:
if is_list_conn:
logger.info(f'连接了第{page_num - 1}页和第{page_num}页的列表段落')
"""接下来可能会漏掉一些特别的一些可以合并的内容，对他们进行段落连接
1. 正文中有时出现一个行顶格，接下来几行缩进的情况。
2. 居中的一些连续单行，如果高度相同，那么可能是一个段落。
"""
for page_num, page in enumerate(pdf_info_dict.values()):
page_paras = page['para_blocks']
new_layout_bbox = new_layout_of_pages[page_num]
__connect_middle_align_text(page_paras, new_layout_bbox, page_num, lang)
__merge_signle_list_text(page_paras, new_layout_bbox, page_num, lang)
# layout展平
for page_num, page in enumerate(pdf_info_dict.values()):
page_paras = page['para_blocks']
page_blocks = [block for layout in page_paras for block in layout]
page['para_blocks'] = page_blocks
magic_pdf/para/para_split_v3.py
View file @ 8afff9ae
......@@ -271,13 +271,18 @@ def __merge_2_text_blocks(block1, block2):
first_span = first_line['spans'][0]
if len(first_span['content']) > 0:
span_start_with_num = first_span['content'][0].isdigit()
span_start_with_big_char = first_span['content'][0].isupper()
if (
abs(block2['bbox_fs'][2] - last_line['bbox'][2])
< line_height
# 上一个block的最后一个line的右边界和block的右边界差距不超过line_height
abs(block2['bbox_fs'][2] - last_line['bbox'][2]) < line_height
# 上一个block的最后一个span不是以特定符号结尾
and not last_span['content'].endswith(LINE_STOP_FLAG)
# 两个block宽度差距超过2倍也不合并
and abs(block1_weight - block2_weight) < min_block_weight
# 下一个block的第一个字符是数字
and not span_start_with_num
# 下一个block的第一个字符是大写字母
and not span_start_with_big_char
):
if block1['page_num'] != block2['page_num']:
for line in block1['lines']:
......
magic_pdf/para/raw_processor.py deleted 100644 → 0
View file @ 4df1eb74
class RawBlockProcessor:
def __init__(self) -> None:
self.y_tolerance = 2
self.pdf_dic = {}
def __span_flags_decomposer(self, span_flags):
"""
Make font flags human readable.
Parameters
----------
self : object
The instance of the class.
span_flags : int
span flags
Returns
-------
l : dict
decomposed flags
"""
l = {
"is_superscript": False,
"is_italic": False,
"is_serifed": False,
"is_sans_serifed": False,
"is_monospaced": False,
"is_proportional": False,
"is_bold": False,
}
if span_flags & 2**0:
l["is_superscript"] = True # 表示上标
if span_flags & 2**1:
l["is_italic"] = True # 表示斜体
if span_flags & 2**2:
l["is_serifed"] = True # 表示衬线字体
else:
l["is_sans_serifed"] = True # 表示非衬线字体
if span_flags & 2**3:
l["is_monospaced"] = True # 表示等宽字体
else:
l["is_proportional"] = True # 表示比例字体
if span_flags & 2**4:
l["is_bold"] = True # 表示粗体
return l
def __make_new_lines(self, raw_lines):
"""
This function makes new lines.
Parameters
----------
self : object
The instance of the class.
raw_lines : list
raw lines
Returns
-------
new_lines : list
new lines
"""
new_lines = []
new_line = None
for raw_line in raw_lines:
raw_line_bbox = raw_line["bbox"]
raw_line_spans = raw_line["spans"]
raw_line_text = "".join([span["text"] for span in raw_line_spans])
raw_line_dir = raw_line.get("dir", None)
decomposed_line_spans = []
for span in raw_line_spans:
raw_flags = span["flags"]
decomposed_flags = self.__span_flags_decomposer(raw_flags)
span["decomposed_flags"] = decomposed_flags
decomposed_line_spans.append(span)
if new_line is None:
new_line = {
"bbox": raw_line_bbox,
"text": raw_line_text,
"dir": raw_line_dir if raw_line_dir else (0, 0),
"spans": decomposed_line_spans,
}
else:
if (
abs(raw_line_bbox[1] - new_line["bbox"][1]) <= self.y_tolerance
and abs(raw_line_bbox[3] - new_line["bbox"][3]) <= self.y_tolerance
):
new_line["bbox"] = (
min(new_line["bbox"][0], raw_line_bbox[0]), # left
new_line["bbox"][1], # top
max(new_line["bbox"][2], raw_line_bbox[2]), # right
raw_line_bbox[3], # bottom
)
new_line["text"] += " " + raw_line_text
new_line["spans"].extend(raw_line_spans)
new_line["dir"] = (
new_line["dir"][0] + raw_line_dir[0],
new_line["dir"][1] + raw_line_dir[1],
)
else:
new_lines.append(new_line)
new_line = {
"bbox": raw_line_bbox,
"text": raw_line_text,
"dir": raw_line_dir if raw_line_dir else (0, 0),
"spans": raw_line_spans,
}
if new_line:
new_lines.append(new_line)
return new_lines
def __make_new_block(self, raw_block):
"""
This function makes a new block.
Parameters
----------
self : object
The instance of the class.
----------
raw_block : dict
a raw block
Returns
-------
new_block : dict
Schema of new_block:
{
"block_id": "block_1",
"bbox": [0, 0, 100, 100],
"text": "This is a block.",
"lines": [
{
"bbox": [0, 0, 100, 100],
"text": "This is a line.",
"spans": [
{
"text": "This is a span.",
"font": "Times New Roman",
"size": 12,
"color": "#000000",
}
],
}
],
}
"""
new_block = {}
block_id = raw_block["number"]
block_bbox = raw_block["bbox"]
block_text = " ".join(span["text"] for line in raw_block["lines"] for span in line["spans"])
raw_lines = raw_block["lines"]
block_lines = self.__make_new_lines(raw_lines)
new_block["block_id"] = block_id
new_block["bbox"] = block_bbox
new_block["text"] = block_text
new_block["lines"] = block_lines
return new_block
def batch_process_blocks(self, pdf_dic):
"""
This function processes the blocks in batch.
Parameters
----------
self : object
The instance of the class.
----------
blocks : list
Input block is a list of raw blocks. Schema can refer to the value of key ""preproc_blocks", demo file is app/pdf_toolbox/tests/preproc_2_parasplit_example.json.
Returns
-------
result_dict : dict
result dictionary
"""
for page_id, blocks in pdf_dic.items():
if page_id.startswith("page_"):
para_blocks = []
if "preproc_blocks" in blocks.keys():
input_blocks = blocks["preproc_blocks"]
for raw_block in input_blocks:
new_block = self.__make_new_block(raw_block)
para_blocks.append(new_block)
blocks["para_blocks"] = para_blocks
return pdf_dic
magic_pdf/para/stats.py deleted 100644 → 0
View file @ 4df1eb74
from collections import Counter
import numpy as np
from magic_pdf.para.commons import *
if sys.version_info[0] >= 3:
sys.stdout.reconfigure(encoding="utf-8") # type: ignore
class BlockStatisticsCalculator:
def __init__(self) -> None:
pass
def __calc_stats_of_new_lines(self, new_lines):
"""
This function calculates the paragraph metrics
Parameters
----------
combined_lines : list
combined lines
Returns
-------
X0 : float
Median of x0 values, which represents the left average boundary of the block
X1 : float
Median of x1 values, which represents the right average boundary of the block
avg_char_width : float
Average of char widths, which represents the average char width of the block
avg_char_height : float
Average of line heights, which represents the average line height of the block
"""
x0_values = []
x1_values = []
char_widths = []
char_heights = []
block_font_types = []
block_font_sizes = []
block_directions = []
if len(new_lines) > 0:
for i, line in enumerate(new_lines):
line_bbox = line["bbox"]
line_text = line["text"]
line_spans = line["spans"]
num_chars = len([ch for ch in line_text if not ch.isspace()])
x0_values.append(line_bbox[0])
x1_values.append(line_bbox[2])
if num_chars > 0:
char_width = (line_bbox[2] - line_bbox[0]) / num_chars
char_widths.append(char_width)
for span in line_spans:
block_font_types.append(span["font"])
block_font_sizes.append(span["size"])
if "dir" in line:
block_directions.append(line["dir"])
# line_font_types = [span["font"] for span in line_spans]
char_heights = [span["size"] for span in line_spans]
X0 = np.median(x0_values) if x0_values else 0
X1 = np.median(x1_values) if x1_values else 0
avg_char_width = sum(char_widths) / len(char_widths) if char_widths else 0
avg_char_height = sum(char_heights) / len(char_heights) if char_heights else 0
# max_freq_font_type = max(set(block_font_types), key=block_font_types.count) if block_font_types else None
max_span_length = 0
max_span_font_type = None
for line in new_lines:
line_spans = line["spans"]
for span in line_spans:
span_length = span["bbox"][2] - span["bbox"][0]
if span_length > max_span_length:
max_span_length = span_length
max_span_font_type = span["font"]
max_freq_font_type = max_span_font_type
avg_font_size = sum(block_font_sizes) / len(block_font_sizes) if block_font_sizes else None
avg_dir_horizontal = sum([dir[0] for dir in block_directions]) / len(block_directions) if block_directions else 0
avg_dir_vertical = sum([dir[1] for dir in block_directions]) / len(block_directions) if block_directions else 0
median_font_size = float(np.median(block_font_sizes)) if block_font_sizes else None
return (
X0,
X1,
avg_char_width,
avg_char_height,
max_freq_font_type,
avg_font_size,
(avg_dir_horizontal, avg_dir_vertical),
median_font_size,
)
def __make_new_block(self, input_block):
new_block = {}
raw_lines = input_block["lines"]
stats = self.__calc_stats_of_new_lines(raw_lines)
block_id = input_block["block_id"]
block_bbox = input_block["bbox"]
block_text = input_block["text"]
block_lines = raw_lines
block_avg_left_boundary = stats[0]
block_avg_right_boundary = stats[1]
block_avg_char_width = stats[2]
block_avg_char_height = stats[3]
block_font_type = stats[4]
block_font_size = stats[5]
block_direction = stats[6]
block_median_font_size = stats[7]
new_block["block_id"] = block_id
new_block["bbox"] = block_bbox
new_block["text"] = block_text
new_block["dir"] = block_direction
new_block["X0"] = block_avg_left_boundary
new_block["X1"] = block_avg_right_boundary
new_block["avg_char_width"] = block_avg_char_width
new_block["avg_char_height"] = block_avg_char_height
new_block["block_font_type"] = block_font_type
new_block["block_font_size"] = block_font_size
new_block["lines"] = block_lines
new_block["median_font_size"] = block_median_font_size
return new_block
def batch_process_blocks(self, pdf_dic):
"""
This function processes the blocks in batch.
Parameters
----------
self : object
The instance of the class.
----------
blocks : list
Input block is a list of raw blocks. Schema can refer to the value of key ""preproc_blocks", demo file is app/pdf_toolbox/tests/preproc_2_parasplit_example.json
Returns
-------
result_dict : dict
result dictionary
"""
for page_id, blocks in pdf_dic.items():
if page_id.startswith("page_"):
para_blocks = []
if "para_blocks" in blocks.keys():
input_blocks = blocks["para_blocks"]
for input_block in input_blocks:
new_block = self.__make_new_block(input_block)
para_blocks.append(new_block)
blocks["para_blocks"] = para_blocks
return pdf_dic
class DocStatisticsCalculator:
def __init__(self) -> None:
pass
def calc_stats_of_doc(self, pdf_dict):
"""
This function computes the statistics of the document
Parameters
----------
result_dict : dict
result dictionary
Returns
-------
statistics : dict
statistics of the document
"""
total_text_length = 0
total_num_blocks = 0
for page_id, blocks in pdf_dict.items():
if page_id.startswith("page_"):
if "para_blocks" in blocks.keys():
para_blocks = blocks["para_blocks"]
for para_block in para_blocks:
total_text_length += len(para_block["text"])
total_num_blocks += 1
avg_text_length = total_text_length / total_num_blocks if total_num_blocks else 0
font_list = []
for page_id, blocks in pdf_dict.items():
if page_id.startswith("page_"):
if "para_blocks" in blocks.keys():
input_blocks = blocks["para_blocks"]
for input_block in input_blocks:
block_text_length = len(input_block.get("text", ""))
if block_text_length < avg_text_length * 0.5:
continue
block_font_type = safe_get(input_block, "block_font_type", "")
block_font_size = safe_get(input_block, "block_font_size", 0)
font_list.append((block_font_type, block_font_size))
font_counter = Counter(font_list)
most_common_font = font_counter.most_common(1)[0] if font_list else (("", 0), 0)
second_most_common_font = font_counter.most_common(2)[1] if len(font_counter) > 1 else (("", 0), 0)
statistics = {
"num_pages": 0,
"num_blocks": 0,
"num_paras": 0,
"num_titles": 0,
"num_header_blocks": 0,
"num_footer_blocks": 0,
"num_watermark_blocks": 0,
"num_vertical_margin_note_blocks": 0,
"most_common_font_type": most_common_font[0][0],
"most_common_font_size": most_common_font[0][1],
"number_of_most_common_font": most_common_font[1],
"second_most_common_font_type": second_most_common_font[0][0],
"second_most_common_font_size": second_most_common_font[0][1],
"number_of_second_most_common_font": second_most_common_font[1],
"avg_text_length": avg_text_length,
}
for page_id, blocks in pdf_dict.items():
if page_id.startswith("page_"):
blocks = pdf_dict[page_id]["para_blocks"]
statistics["num_pages"] += 1
for block_id, block_data in enumerate(blocks):
statistics["num_blocks"] += 1
if "paras" in block_data.keys():
statistics["num_paras"] += len(block_data["paras"])
for line in block_data["lines"]:
if line.get("is_title", 0):
statistics["num_titles"] += 1
if block_data.get("is_header", 0):
statistics["num_header_blocks"] += 1
if block_data.get("is_footer", 0):
statistics["num_footer_blocks"] += 1
if block_data.get("is_watermark", 0):
statistics["num_watermark_blocks"] += 1
if block_data.get("is_vertical_margin_note", 0):
statistics["num_vertical_margin_note_blocks"] += 1
pdf_dict["statistics"] = statistics
return pdf_dict
magic_pdf/para/title_processor.py deleted 100644 → 0
View file @ 4df1eb74
import os
import re
import numpy as np
from magic_pdf.libs.nlp_utils import NLPModels
from magic_pdf.para.commons import *
if sys.version_info[0] >= 3:
sys.stdout.reconfigure(encoding="utf-8") # type: ignore
class TitleProcessor:
def __init__(self, *doc_statistics) -> None:
if len(doc_statistics) > 0:
self.doc_statistics = doc_statistics[0]
self.nlp_model = NLPModels()
self.MAX_TITLE_LEVEL = 3
self.numbered_title_pattern = r"""
^ # 行首
( # 开始捕获组
[\(\（]\d+[\)\）] # 括号内数字，支持中文和英文括号，例如：(1) 或 （1）
|\d+[\)\）]\s # 数字后跟右括号和空格，支持中文和英文括号，例如：2) 或 2）
|[\(\（][A-Z][\)\）] # 括号内大写字母，支持中文和英文括号，例如：(A) 或 （A）
|[A-Z][\)\）]\s # 大写字母后跟右括号和空格，例如：A) 或 A）
|[\(\（][IVXLCDM]+[\)\）] # 括号内罗马数字，支持中文和英文括号，例如：(I) 或 （I）
|[IVXLCDM]+[\)\）]\s # 罗马数字后跟右括号和空格，例如：I) 或 I）
|\d+(\.\d+)*\s # 数字或复合数字编号后跟空格，例如：1. 或 3.2.1
|[一二三四五六七八九十百千]+[、\s] # 中文序号后跟顿号和空格，例如：一、
|[\（|\(][一二三四五六七八九十百千]+[\）|\)]\s* # 中文括号内中文序号后跟空格，例如：（一）
|[A-Z]\.\d+(\.\d+)?\s # 大写字母后跟点和数字，例如：A.1 或 A.1.1
|[\(\（][a-z][\)\）] # 括号内小写字母，支持中文和英文括号，例如：(a) 或 （a）
|[a-z]\)\s # 小写字母后跟右括号和空格，例如：a)
|[A-Z]-\s # 大写字母后跟短横线和空格，例如：A-
|\w+:\s # 英文序号词后跟冒号和空格，例如：First:
|第[一二三四五六七八九十百千]+[章节部分条款]\s # 以“第”开头的中文标题后跟空格
|[IVXLCDM]+\. # 罗马数字后跟点，例如：I.
|\d+\.\s # 单个数字后跟点和空格，例如：1.
) # 结束捕获组
.+ # 标题的其余部分
"""
def _is_potential_title(
self,
curr_line,
prev_line,
prev_line_is_title,
next_line,
avg_char_width,
avg_char_height,
median_font_size,
):
"""
This function checks if the line is a potential title.
Parameters
----------
curr_line : dict
current line
prev_line : dict
previous line
next_line : dict
next line
avg_char_width : float
average of char widths
avg_char_height : float
average of line heights
Returns
-------
bool
True if the line is a potential title, False otherwise.
"""
def __is_line_centered(line_bbox, page_bbox, avg_char_width):
"""
This function checks if the line is centered on the page
Parameters
----------
line_bbox : list
bbox of the line
page_bbox : list
bbox of the page
avg_char_width : float
average of char widths
Returns
-------
bool
True if the line is centered on the page, False otherwise.
"""
horizontal_ratio = 0.5
horizontal_thres = horizontal_ratio * avg_char_width
x0, _, x1, _ = line_bbox
_, _, page_x1, _ = page_bbox
return abs((x0 + x1) / 2 - page_x1 / 2) < horizontal_thres
def __is_bold_font_line(line):
"""
Check if a line contains any bold font style.
"""
def _is_bold_span(span):
# if span text is empty or only contains space, return False
if not span["text"].strip():
return False
return bool(span["flags"] & 2**4) # Check if the font is bold
for span in line["spans"]:
if not _is_bold_span(span):
return False
return True
def __is_italic_font_line(line):
"""
Check if a line contains any italic font style.
"""
def __is_italic_span(span):
return bool(span["flags"] & 2**1) # Check if the font is italic
for span in line["spans"]:
if not __is_italic_span(span):
return False
return True
def __is_punctuation_heavy(line_text):
"""
Check if the line contains a high ratio of punctuation marks, which may indicate
that the line is not a title.
Parameters:
line_text (str): Text of the line.
Returns:
bool: True if the line is heavy with punctuation, False otherwise.
"""
# Pattern for common title format like "X.Y. Title"
pattern = r"\b\d+\.\d+\..*\b"
# If the line matches the title format, return False
if re.match(pattern, line_text.strip()):
return False
# Find all punctuation marks in the line
punctuation_marks = re.findall(r"[^\w\s]", line_text)
number_of_punctuation_marks = len(punctuation_marks)
text_length = len(line_text)
if text_length == 0:
return False
punctuation_ratio = number_of_punctuation_marks / text_length
if punctuation_ratio >= 0.1:
return True
return False
def __has_mixed_font_styles(spans, strict_mode=False):
"""
This function checks if the line has mixed font styles, the strict mode will compare the font types
Parameters
----------
spans : list
spans of the line
strict_mode : bool
True for strict mode, the font types will be fully compared
False for non-strict mode, the font types will be compared by the most longest common prefix
Returns
-------
bool
True if the line has mixed font styles, False otherwise.
"""
if strict_mode:
font_styles = set()
for span in spans:
font_style = span["font"].lower()
font_styles.add(font_style)
return len(font_styles) > 1
else: # non-strict mode
font_styles = []
for span in spans:
font_style = span["font"].lower()
font_styles.append(font_style)
if len(font_styles) > 1:
longest_common_prefix = os.path.commonprefix(font_styles)
if len(longest_common_prefix) > 0:
return False
else:
return True
else:
return False
def __is_different_font_type_from_neighbors(curr_line_font_type, prev_line_font_type, next_line_font_type):
"""
This function checks if the current line has a different font type from the previous and next lines
Parameters
----------
curr_line_font_type : str
font type of the current line
prev_line_font_type : str
font type of the previous line
next_line_font_type : str
font type of the next line
Returns
-------
bool
True if the current line has a different font type from the previous and next lines, False otherwise.
"""
return all(
curr_line_font_type != other_font_type.lower()
for other_font_type in [prev_line_font_type, next_line_font_type]
if other_font_type is not None
)
def __is_larger_font_size_from_neighbors(curr_line_font_size, prev_line_font_size, next_line_font_size):
"""
This function checks if the current line has a larger font size than the previous and next lines
Parameters
----------
curr_line_font_size : float
font size of the current line
prev_line_font_size : float
font size of the previous line
next_line_font_size : float
font size of the next line
Returns
-------
bool
True if the current line has a larger font size than the previous and next lines, False otherwise.
"""
return all(
curr_line_font_size > other_font_size * 1.2
for other_font_size in [prev_line_font_size, next_line_font_size]
if other_font_size is not None
)
def __is_similar_to_pre_line(curr_line_font_type, prev_line_font_type, curr_line_font_size, prev_line_font_size):
"""
This function checks if the current line is similar to the previous line
Parameters
----------
curr_line : dict
current line
prev_line : dict
previous line
Returns
-------
bool
True if the current line is similar to the previous line, False otherwise.
"""
if curr_line_font_type == prev_line_font_type and curr_line_font_size == prev_line_font_size:
return True
else:
return False
def __is_same_font_type_of_docAvg(curr_line_font_type):
"""
This function checks if the current line has the same font type as the document average font type
Parameters
----------
curr_line_font_type : str
font type of the current line
Returns
-------
bool
True if the current line has the same font type as the document average font type, False otherwise.
"""
doc_most_common_font_type = safe_get(self.doc_statistics, "most_common_font_type", "").lower()
doc_second_most_common_font_type = safe_get(self.doc_statistics, "second_most_common_font_type", "").lower()
return curr_line_font_type.lower() in [doc_most_common_font_type, doc_second_most_common_font_type]
def __is_font_size_not_less_than_docAvg(curr_line_font_size, ratio: float = 1):
"""
This function checks if the current line has a large enough font size
Parameters
----------
curr_line_font_size : float
font size of the current line
ratio : float
ratio of the current line font size to the document average font size
Returns
-------
bool
True if the current line has a large enough font size, False otherwise.
"""
doc_most_common_font_size = safe_get(self.doc_statistics, "most_common_font_size", 0)
doc_second_most_common_font_size = safe_get(self.doc_statistics, "second_most_common_font_size", 0)
doc_avg_font_size = min(doc_most_common_font_size, doc_second_most_common_font_size)
return curr_line_font_size >= doc_avg_font_size * ratio
def __is_sufficient_spacing_above_and_below(
curr_line_bbox,
prev_line_bbox,
next_line_bbox,
avg_char_height,
median_font_size,
):
"""
This function checks if the current line has sufficient spacing above and below
Parameters
----------
curr_line_bbox : list
bbox of the current line
prev_line_bbox : list
bbox of the previous line
next_line_bbox : list
bbox of the next line
avg_char_width : float
average of char widths
avg_char_height : float
average of line heights
Returns
-------
bool
True if the current line has sufficient spacing above and below, False otherwise.
"""
vertical_ratio = 1.25
vertical_thres = vertical_ratio * median_font_size
_, y0, _, y1 = curr_line_bbox
sufficient_spacing_above = False
if prev_line_bbox:
vertical_spacing_above = min(y0 - prev_line_bbox[1], y1 - prev_line_bbox[3])
sufficient_spacing_above = vertical_spacing_above > vertical_thres
else:
sufficient_spacing_above = True
sufficient_spacing_below = False
if next_line_bbox:
vertical_spacing_below = min(next_line_bbox[1] - y0, next_line_bbox[3] - y1)
sufficient_spacing_below = vertical_spacing_below > vertical_thres
else:
sufficient_spacing_below = True
return (sufficient_spacing_above, sufficient_spacing_below)
def __is_word_list_line_by_rules(curr_line_text):
"""
This function checks if the current line is a word list
Parameters
----------
curr_line_text : str
text of the current line
Returns
-------
bool
True if the current line is a name list, False otherwise.
"""
# name_list_pattern = r"([a-zA-Z][a-zA-Z\s]{0,20}[a-zA-Z]|[\u4e00-\u9fa5·]{2,16})(?=[，,;；\s]|$)"
name_list_pattern = r"(?<![\u4e00-\u9fa5])([A-Z][a-z]{0,19}\s[A-Z][a-z]{0,19}|[\u4e00-\u9fa5]{2,6})(?=[，,;；\s]|$)"
compiled_pattern = re.compile(name_list_pattern)
if compiled_pattern.search(curr_line_text):
return True
else:
return False
# """
def __get_text_catgr_by_nlp(curr_line_text):
"""
This function checks if the current line is a name list using nlp model, such as spacy
Parameters
----------
curr_line_text : str
text of the current line
Returns
-------
bool
True if the current line is a name list, False otherwise.
"""
result = self.nlp_model.detect_entity_catgr_using_nlp(curr_line_text)
return result
# """
def __is_numbered_title(curr_line_text):
"""
This function checks if the current line is a numbered list
Parameters
----------
curr_line_text : str
text of the current line
Returns
-------
bool
True if the current line is a numbered list, False otherwise.
"""
compiled_pattern = re.compile(self.numbered_title_pattern, re.VERBOSE)
if compiled_pattern.search(curr_line_text):
return True
else:
return False
def __is_end_with_ending_puncs(line_text):
"""
This function checks if the current line ends with a ending punctuation mark
Parameters
----------
line_text : str
text of the current line
Returns
-------
bool
True if the current line ends with a punctuation mark, False otherwise.
"""
end_puncs = [".", "?", "!", "。", "？", "！", "…"]
line_text = line_text.rstrip()
if line_text[-1] in end_puncs:
return True
return False
def __contains_only_no_meaning_symbols(line_text):
"""
This function checks if the current line contains only symbols that have no meaning, if so, it is not a title.
Situation contains:
1. Only have punctuation marks
2. Only have other non-meaning symbols
Parameters
----------
line_text : str
text of the current line
Returns
-------
bool
True if the current line contains only symbols that have no meaning, False otherwise.
"""
punctuation_marks = re.findall(r"[^\w\s]", line_text) # find all punctuation marks
number_of_punctuation_marks = len(punctuation_marks)
text_length = len(line_text)
if text_length == 0:
return False
punctuation_ratio = number_of_punctuation_marks / text_length
if punctuation_ratio >= 0.9:
return True
return False
def __is_equation(line_text):
"""
This function checks if the current line is an equation.
Parameters
----------
line_text : str
Returns
-------
bool
True if the current line is an equation, False otherwise.
"""
equation_reg = r"\$.*?\\overline.*?\$" # to match interline equations
if re.search(equation_reg, line_text):
return True
else:
return False
def __is_title_by_len(text, max_length=200):
"""
This function checks if the current line is a title by length.
Parameters
----------
text : str
text of the current line
max_length : int
max length of the title
Returns
-------
bool
True if the current line is a title, False otherwise.
"""
text = text.strip()
return len(text) <= max_length
def __compute_line_font_type_and_size(curr_line):
"""
This function computes the font type and font size of the line.
Parameters
----------
line : dict
line
Returns
-------
font_type : str
font type of the line
font_size : float
font size of the line
"""
spans = curr_line["spans"]
max_accumulated_length = 0
max_span_font_size = curr_line["spans"][0]["size"] # default value, float type
max_span_font_type = curr_line["spans"][0]["font"].lower() # default value, string type
for span in spans:
if span["text"].isspace():
continue
span_length = span["bbox"][2] - span["bbox"][0]
if span_length > max_accumulated_length:
max_accumulated_length = span_length
max_span_font_size = span["size"]
max_span_font_type = span["font"].lower()
return max_span_font_type, max_span_font_size
"""
Title detecting main Process.
"""
"""
Basic features about the current line.
"""
curr_line_bbox = curr_line["bbox"]
curr_line_text = curr_line["text"]
curr_line_font_type, curr_line_font_size = __compute_line_font_type_and_size(curr_line)
if len(curr_line_text.strip()) == 0: # skip empty lines
return False
prev_line_bbox = prev_line["bbox"] if prev_line else None
if prev_line:
prev_line_font_type, prev_line_font_size = __compute_line_font_type_and_size(prev_line)
else:
prev_line_font_type, prev_line_font_size = None, None
next_line_bbox = next_line["bbox"] if next_line else None
if next_line:
next_line_font_type, next_line_font_size = __compute_line_font_type_and_size(next_line)
else:
next_line_font_type, next_line_font_size = None, None
"""
Aggregated features about the current line.
"""
is_italc_font = __is_italic_font_line(curr_line)
is_bold_font = __is_bold_font_line(curr_line)
is_font_size_little_less_than_doc_avg = __is_font_size_not_less_than_docAvg(curr_line_font_size, ratio=0.8)
is_font_size_not_less_than_doc_avg = __is_font_size_not_less_than_docAvg(curr_line_font_size, ratio=1)
is_much_larger_font_than_doc_avg = __is_font_size_not_less_than_docAvg(curr_line_font_size, ratio=1.6)
is_not_same_font_type_of_docAvg = not __is_same_font_type_of_docAvg(curr_line_font_type)
is_potential_title_font = is_bold_font or is_font_size_not_less_than_doc_avg or is_not_same_font_type_of_docAvg
is_mix_font_styles_strict = __has_mixed_font_styles(curr_line["spans"], strict_mode=True)
is_mix_font_styles_loose = __has_mixed_font_styles(curr_line["spans"], strict_mode=False)
is_punctuation_heavy = __is_punctuation_heavy(curr_line_text)
is_word_list_line_by_rules = __is_word_list_line_by_rules(curr_line_text)
is_person_or_org_list_line_by_nlp = __get_text_catgr_by_nlp(curr_line_text) in ["PERSON", "GPE", "ORG"]
is_font_size_larger_than_neighbors = __is_larger_font_size_from_neighbors(
curr_line_font_size, prev_line_font_size, next_line_font_size
)
is_font_type_diff_from_neighbors = __is_different_font_type_from_neighbors(
curr_line_font_type, prev_line_font_type, next_line_font_type
)
has_sufficient_spaces_above, has_sufficient_spaces_below = __is_sufficient_spacing_above_and_below(
curr_line_bbox, prev_line_bbox, next_line_bbox, avg_char_height, median_font_size
)
is_similar_to_pre_line = __is_similar_to_pre_line(
curr_line_font_type, prev_line_font_type, curr_line_font_size, prev_line_font_size
)
"""
Further aggregated features about the current line.
Attention:
Features that start with __ are for internal use.
"""
__is_line_left_aligned_from_neighbors = is_line_left_aligned_from_neighbors(
curr_line_bbox, prev_line_bbox, next_line_bbox, avg_char_width
)
__is_font_diff_from_neighbors = is_font_size_larger_than_neighbors or is_font_type_diff_from_neighbors
is_a_left_inline_title = (
is_mix_font_styles_strict and __is_line_left_aligned_from_neighbors and __is_font_diff_from_neighbors
)
is_title_by_check_prev_line = prev_line is None and has_sufficient_spaces_above and is_potential_title_font
is_title_by_check_next_line = next_line is None and has_sufficient_spaces_below and is_potential_title_font
is_title_by_check_pre_and_next_line = (
(prev_line is not None or next_line is not None)
and has_sufficient_spaces_above
and has_sufficient_spaces_below
and is_potential_title_font
)
is_numbered_title = __is_numbered_title(curr_line_text) and (
(has_sufficient_spaces_above or prev_line is None) and (has_sufficient_spaces_below or next_line is None)
)
is_not_end_with_ending_puncs = not __is_end_with_ending_puncs(curr_line_text)
is_not_only_no_meaning_symbols = not __contains_only_no_meaning_symbols(curr_line_text)
is_equation = __is_equation(curr_line_text)
is_title_by_len = __is_title_by_len(curr_line_text)
"""
Decide if the line is a title.
"""
# is_title = False
# if prev_line_is_title:
is_title = (
is_not_end_with_ending_puncs # not end with ending punctuation marks
and is_not_only_no_meaning_symbols # not only have no meaning symbols
and is_title_by_len # is a title by length, default max length is 200
and not is_equation # an interline equation should never be a title
and is_potential_title_font # is a potential title font, which is bold or larger than the document average font size or not the same font type as the document average font type
and (
(is_not_same_font_type_of_docAvg and is_font_size_not_less_than_doc_avg)
or (is_bold_font and is_much_larger_font_than_doc_avg and is_not_same_font_type_of_docAvg)
or (
is_much_larger_font_than_doc_avg
and (is_title_by_check_prev_line or is_title_by_check_next_line or is_title_by_check_pre_and_next_line)
)
or (
is_font_size_little_less_than_doc_avg
and is_bold_font
and (is_title_by_check_prev_line or is_title_by_check_next_line or is_title_by_check_pre_and_next_line)
)
) # not the same font type as the document average font type, which includes the most common font type and the second most common font type
and (
(
not is_person_or_org_list_line_by_nlp
and (
is_much_larger_font_than_doc_avg
or (is_not_same_font_type_of_docAvg and is_font_size_not_less_than_doc_avg)
)
)
or (
not (is_word_list_line_by_rules and is_person_or_org_list_line_by_nlp)
and not is_a_left_inline_title
and not is_punctuation_heavy
and (is_title_by_check_prev_line or is_title_by_check_next_line or is_title_by_check_pre_and_next_line)
)
or (
is_person_or_org_list_line_by_nlp
and (is_bold_font and is_much_larger_font_than_doc_avg and is_not_same_font_type_of_docAvg)
and (is_bold_font and is_much_larger_font_than_doc_avg and is_not_same_font_type_of_docAvg)
)
or (is_numbered_title and not is_a_left_inline_title)
)
)
# ) or (is_similar_to_pre_line and prev_line_is_title)
is_name_or_org_list_to_be_removed = (
(is_person_or_org_list_line_by_nlp)
and is_punctuation_heavy
and (is_title_by_check_prev_line or is_title_by_check_next_line or is_title_by_check_pre_and_next_line)
) and not is_title
if is_name_or_org_list_to_be_removed:
is_author_or_org_list = True
# print curr_line_text to check
# print_yellow(f"Text of is_author_or_org_list: {curr_line_text}")
else:
is_author_or_org_list = False
"""
# print reason why the line is a title
if is_title:
print_green("This line is a title.")
print_green("↓" * 10)
print()
print("curr_line_text: ", curr_line_text)
print()
# print reason why the line is not a title
line_text = curr_line_text.strip()
test_text = "Career/Personal Life"
text_content_condition = line_text == test_text
if not is_title and text_content_condition: # Print specific line
# if not is_title: # Print each line
print_red("This line is not a title.")
print_red("↓" * 10)
print()
print("curr_line_text: ", curr_line_text)
print()
if is_not_end_with_ending_puncs:
print_green(f"is_not_end_with_ending_puncs")
else:
print_red(f"is_end_with_ending_puncs")
if is_not_only_no_meaning_symbols:
print_green(f"is_not_only_no_meaning_symbols")
else:
print_red(f"is_only_no_meaning_symbols")
if is_title_by_len:
print_green(f"is_title_by_len: {is_title_by_len}")
else:
print_red(f"is_not_title_by_len: {is_title_by_len}")
if is_equation:
print_red(f"is_equation")
else:
print_green(f"is_not_equation")
if is_potential_title_font:
print_green(f"is_potential_title_font")
else:
print_red(f"is_not_potential_title_font")
if is_punctuation_heavy:
print_red("is_punctuation_heavy")
else:
print_green("is_not_punctuation_heavy")
if is_bold_font:
print_green(f"is_bold_font")
else:
print_red(f"is_not_bold_font")
if is_font_size_not_less_than_doc_avg:
print_green(f"is_larger_font_than_doc_avg")
else:
print_red(f"is_not_larger_font_than_doc_avg")
if is_much_larger_font_than_doc_avg:
print_green(f"is_much_larger_font_than_doc_avg")
else:
print_red(f"is_not_much_larger_font_than_doc_avg")
if is_not_same_font_type_of_docAvg:
print_green(f"is_not_same_font_type_of_docAvg")
else:
print_red(f"is_same_font_type_of_docAvg")
if is_word_list_line_by_rules:
print_red("is_word_list_line_by_rules")
else:
print_green("is_not_name_list_by_rules")
if is_person_or_org_list_line_by_nlp:
print_red("is_person_or_org_list_line_by_nlp")
else:
print_green("is_not_person_or_org_list_line_by_nlp")
if not is_numbered_title:
print_red("is_not_numbered_title")
else:
print_green("is_numbered_title")
if is_a_left_inline_title:
print_red("is_a_left_inline_title")
else:
print_green("is_not_a_left_inline_title")
if not is_title_by_check_prev_line:
print_red("is_not_title_by_check_prev_line")
else:
print_green("is_title_by_check_prev_line")
if not is_title_by_check_next_line:
print_red("is_not_title_by_check_next_line")
else:
print_green("is_title_by_check_next_line")
if not is_title_by_check_pre_and_next_line:
print_red("is_not_title_by_check_pre_and_next_line")
else:
print_green("is_title_by_check_pre_and_next_line")
# print_green("Common features:")
# print_green("↓" * 10)
# print(f" curr_line_font_type: {curr_line_font_type}")
# print(f" curr_line_font_size: {curr_line_font_size}")
# print()
"""
return is_title, is_author_or_org_list
def _detect_block_title(self, input_block):
"""
Use the functions 'is_potential_title' to detect titles of each paragraph block.
If a line is a title, then the value of key 'is_title' of the line will be set to True.
"""
raw_lines = input_block["lines"]
prev_line_is_title_flag = False
for i, curr_line in enumerate(raw_lines):
prev_line = raw_lines[i - 1] if i > 0 else None
next_line = raw_lines[i + 1] if i < len(raw_lines) - 1 else None
blk_avg_char_width = input_block["avg_char_width"]
blk_avg_char_height = input_block["avg_char_height"]
blk_media_font_size = input_block["median_font_size"]
is_title, is_author_or_org_list = self._is_potential_title(
curr_line,
prev_line,
prev_line_is_title_flag,
next_line,
blk_avg_char_width,
blk_avg_char_height,
blk_media_font_size,
)
if is_title:
curr_line["is_title"] = is_title
prev_line_is_title_flag = True
else:
curr_line["is_title"] = False
prev_line_is_title_flag = False
if is_author_or_org_list:
curr_line["is_author_or_org_list"] = is_author_or_org_list
else:
curr_line["is_author_or_org_list"] = False
return input_block
def batch_process_blocks_detect_titles(self, pdf_dic):
"""
This function batch process the blocks to detect titles.
Parameters
----------
pdf_dict : dict
result dictionary
Returns
-------
pdf_dict : dict
result dictionary
"""
num_titles = 0
for page_id, blocks in pdf_dic.items():
if page_id.startswith("page_"):
para_blocks = []
if "para_blocks" in blocks.keys():
para_blocks = blocks["para_blocks"]
all_single_line_blocks = []
for block in para_blocks:
if len(block["lines"]) == 1:
all_single_line_blocks.append(block)
new_para_blocks = []
if not len(all_single_line_blocks) == len(para_blocks): # Not all blocks are single line blocks.
for para_block in para_blocks:
new_block = self._detect_block_title(para_block)
new_para_blocks.append(new_block)
num_titles += sum([line.get("is_title", 0) for line in new_block["lines"]])
else: # All blocks are single line blocks.
for para_block in para_blocks:
new_para_blocks.append(para_block)
num_titles += sum([line.get("is_title", 0) for line in para_block["lines"]])
para_blocks = new_para_blocks
blocks["para_blocks"] = para_blocks
for para_block in para_blocks:
all_titles = all(safe_get(line, "is_title", False) for line in para_block["lines"])
para_text_len = sum([len(line["text"]) for line in para_block["lines"]])
if (
all_titles and para_text_len < 200
): # total length of the paragraph is less than 200, more than this should not be a title
para_block["is_block_title"] = 1
else:
para_block["is_block_title"] = 0
all_name_or_org_list_to_be_removed = all(
safe_get(line, "is_author_or_org_list", False) for line in para_block["lines"]
)
if all_name_or_org_list_to_be_removed and page_id == "page_0":
para_block["is_block_an_author_or_org_list"] = 1
else:
para_block["is_block_an_author_or_org_list"] = 0
pdf_dic["statistics"]["num_titles"] = num_titles
return pdf_dic
def __determine_size_based_level(self, title_blocks):
"""
This function determines the title level based on the font size of the title.
Parameters
----------
title_blocks : list
Returns
-------
title_blocks : list
"""
font_sizes = np.array([safe_get(tb["block"], "block_font_size", 0) for tb in title_blocks])
# Use the mean and std of font sizes to remove extreme values
mean_font_size = np.mean(font_sizes)
std_font_size = np.std(font_sizes)
min_extreme_font_size = mean_font_size - std_font_size # type: ignore
max_extreme_font_size = mean_font_size + std_font_size # type: ignore
# Compute the threshold for title level
middle_font_sizes = font_sizes[(font_sizes > min_extreme_font_size) & (font_sizes < max_extreme_font_size)]
if middle_font_sizes.size > 0:
middle_mean_font_size = np.mean(middle_font_sizes)
level_threshold = middle_mean_font_size
else:
level_threshold = mean_font_size
for tb in title_blocks:
title_block = tb["block"]
title_font_size = safe_get(title_block, "block_font_size", 0)
current_level = 1 # Initialize title level, the biggest level is 1
# print(f"Before adjustment by font size, {current_level}")
if title_font_size >= max_extreme_font_size:
current_level = 1
elif title_font_size <= min_extreme_font_size:
current_level = 3
elif float(title_font_size) >= float(level_threshold):
current_level = 2
else:
current_level = 3
# print(f"After adjustment by font size, {current_level}")
title_block["block_title_level"] = current_level
return title_blocks
def batch_process_blocks_recog_title_level(self, pdf_dic):
title_blocks = []
# Collect all titles
for page_id, blocks in pdf_dic.items():
if page_id.startswith("page_"):
para_blocks = blocks.get("para_blocks", [])
for block in para_blocks:
if block.get("is_block_title"):
title_obj = {"page_id": page_id, "block": block}
title_blocks.append(title_obj)
# Determine title level
if title_blocks:
# Determine title level based on font size
title_blocks = self.__determine_size_based_level(title_blocks)
return pdf_dic
magic_pdf/pdf_parse_union_core.py deleted 100644 → 0
View file @ 4df1eb74
import time
from loguru import logger
from magic_pdf.config.drop_reason import DropReason
from magic_pdf.config.ocr_content_type import ContentType
from magic_pdf.layout.layout_sort import (LAYOUT_UNPROC, get_bboxes_layout,
get_columns_cnt_of_layout)
from magic_pdf.libs.commons import fitz, get_delta_time
from magic_pdf.libs.convert_utils import dict_to_list
from magic_pdf.libs.hash_utils import compute_md5
from magic_pdf.libs.local_math import float_equal
from magic_pdf.model.magic_model import MagicModel
from magic_pdf.para.para_split_v2 import para_split
from magic_pdf.pre_proc.citationmarker_remove import remove_citation_marker
from magic_pdf.pre_proc.construct_page_dict import \
ocr_construct_page_component_v2
from magic_pdf.pre_proc.cut_image import ocr_cut_image_and_table
from magic_pdf.pre_proc.equations_replace import (
combine_chars_to_pymudict, remove_chars_in_text_blocks,
replace_equations_in_textblock)
from magic_pdf.pre_proc.ocr_detect_all_bboxes import \
ocr_prepare_bboxes_for_layout_split
from magic_pdf.pre_proc.ocr_dict_merge import (fill_spans_in_blocks,
fix_block_spans,
fix_discarded_block,
sort_blocks_by_layout)
from magic_pdf.pre_proc.ocr_span_list_modify import (
get_qa_need_list_v2, remove_overlaps_low_confidence_spans,
remove_overlaps_min_spans)
from magic_pdf.pre_proc.resolve_bbox_conflict import \
check_useful_block_horizontal_overlap
def remove_horizontal_overlap_block_which_smaller(all_bboxes):
useful_blocks = []
for bbox in all_bboxes:
useful_blocks.append({'bbox': bbox[:4]})
is_useful_block_horz_overlap, smaller_bbox, bigger_bbox = (
check_useful_block_horizontal_overlap(useful_blocks)
)
if is_useful_block_horz_overlap:
logger.warning(
f'skip this page, reason: {DropReason.USEFUL_BLOCK_HOR_OVERLAP}, smaller bbox is {smaller_bbox}, bigger bbox is {bigger_bbox}'
)
for bbox in all_bboxes.copy():
if smaller_bbox == bbox[:4]:
all_bboxes.remove(bbox)
return is_useful_block_horz_overlap, all_bboxes
def __replace_STX_ETX(text_str: str):
"""Replace \u0002 and \u0003, as these characters become garbled when extracted using pymupdf. In fact, they were originally quotation marks.
Drawback: This issue is only observed in English text; it has not been found in Chinese text so far.
Args:
text_str (str): raw text
Returns:
_type_: replaced text
"""
if text_str:
s = text_str.replace('\u0002', "'")
s = s.replace('\u0003', "'")
return s
return text_str
def txt_spans_extract(pdf_page, inline_equations, interline_equations):
text_raw_blocks = pdf_page.get_text('dict', flags=fitz.TEXTFLAGS_TEXT)['blocks']
char_level_text_blocks = pdf_page.get_text('rawdict', flags=fitz.TEXTFLAGS_TEXT)[
'blocks'
]
text_blocks = combine_chars_to_pymudict(text_raw_blocks, char_level_text_blocks)
text_blocks = replace_equations_in_textblock(
text_blocks, inline_equations, interline_equations
)
text_blocks = remove_citation_marker(text_blocks)
text_blocks = remove_chars_in_text_blocks(text_blocks)
spans = []
for v in text_blocks:
for line in v['lines']:
for span in line['spans']:
bbox = span['bbox']
if float_equal(bbox[0], bbox[2]) or float_equal(bbox[1], bbox[3]):
continue
if span.get('type') not in (
ContentType.InlineEquation,
ContentType.InterlineEquation,
):
spans.append(
{
'bbox': list(span['bbox']),
'content': __replace_STX_ETX(span['text']),
'type': ContentType.Text,
'score': 1.0,
}
)
return spans
def replace_text_span(pymu_spans, ocr_spans):
return list(filter(lambda x: x['type'] != ContentType.Text, ocr_spans)) + pymu_spans
def parse_page_core(
pdf_docs, magic_model, page_id, pdf_bytes_md5, imageWriter, parse_mode
):
need_drop = False
drop_reason = []
"""从magic_model对象中获取后面会用到的区块信息"""
img_blocks = magic_model.get_imgs(page_id)
table_blocks = magic_model.get_tables(page_id)
discarded_blocks = magic_model.get_discarded(page_id)
text_blocks = magic_model.get_text_blocks(page_id)
title_blocks = magic_model.get_title_blocks(page_id)
inline_equations, interline_equations, interline_equation_blocks = (
magic_model.get_equations(page_id)
)
page_w, page_h = magic_model.get_page_size(page_id)
spans = magic_model.get_all_spans(page_id)
"""根据parse_mode，构造spans"""
if parse_mode == 'txt':
"""ocr 中文本类的 span 用 pymu spans 替换！"""
pymu_spans = txt_spans_extract(
pdf_docs[page_id], inline_equations, interline_equations
)
spans = replace_text_span(pymu_spans, spans)
elif parse_mode == 'ocr':
pass
else:
raise Exception('parse_mode must be txt or ocr')
"""删除重叠spans中置信度较低的那些"""
spans, dropped_spans_by_confidence = remove_overlaps_low_confidence_spans(spans)
"""删除重叠spans中较小的那些"""
spans, dropped_spans_by_span_overlap = remove_overlaps_min_spans(spans)
"""对image和table截图"""
spans = ocr_cut_image_and_table(
spans, pdf_docs[page_id], page_id, pdf_bytes_md5, imageWriter
)
"""将所有区块的bbox整理到一起"""
# interline_equation_blocks参数不够准，后面切换到interline_equations上
interline_equation_blocks = []
if len(interline_equation_blocks) > 0:
all_bboxes, all_discarded_blocks, drop_reasons = (
ocr_prepare_bboxes_for_layout_split(
img_blocks,
table_blocks,
discarded_blocks,
text_blocks,
title_blocks,
interline_equation_blocks,
page_w,
page_h,
)
)
else:
all_bboxes, all_discarded_blocks, drop_reasons = (
ocr_prepare_bboxes_for_layout_split(
img_blocks,
table_blocks,
discarded_blocks,
text_blocks,
title_blocks,
interline_equations,
page_w,
page_h,
)
)
if len(drop_reasons) > 0:
need_drop = True
drop_reason.append(DropReason.OVERLAP_BLOCKS_CAN_NOT_SEPARATION)
"""先处理不需要排版的discarded_blocks"""
discarded_block_with_spans, spans = fill_spans_in_blocks(
all_discarded_blocks, spans, 0.4
)
fix_discarded_blocks = fix_discarded_block(discarded_block_with_spans)
"""如果当前页面没有bbox则跳过"""
if len(all_bboxes) == 0:
logger.warning(f'skip this page, not found useful bbox, page_id: {page_id}')
return ocr_construct_page_component_v2(
[],
[],
page_id,
page_w,
page_h,
[],
[],
[],
interline_equations,
fix_discarded_blocks,
need_drop,
drop_reason,
)
"""在切分之前，先检查一下bbox是否有左右重叠的情况，如果有，那么就认为这个pdf暂时没有能力处理好，这种左右重叠的情况大概率是由于pdf里的行间公式、表格没有被正确识别出来造成的 """
while True: # 循环检查左右重叠的情况，如果存在就删除掉较小的那个bbox，直到不存在左右重叠的情况
is_useful_block_horz_overlap, all_bboxes = (
remove_horizontal_overlap_block_which_smaller(all_bboxes)
)
if is_useful_block_horz_overlap:
need_drop = True
drop_reason.append(DropReason.USEFUL_BLOCK_HOR_OVERLAP)
else:
break
"""根据区块信息计算layout"""
page_boundry = [0, 0, page_w, page_h]
layout_bboxes, layout_tree = get_bboxes_layout(all_bboxes, page_boundry, page_id)
if len(text_blocks) > 0 and len(all_bboxes) > 0 and len(layout_bboxes) == 0:
logger.warning(
f'skip this page, page_id: {page_id}, reason: {DropReason.CAN_NOT_DETECT_PAGE_LAYOUT}'
)
need_drop = True
drop_reason.append(DropReason.CAN_NOT_DETECT_PAGE_LAYOUT)
"""以下去掉复杂的布局和超过2列的布局"""
if any(
[lay['layout_label'] == LAYOUT_UNPROC for lay in layout_bboxes]
): # 复杂的布局
logger.warning(
f'skip this page, page_id: {page_id}, reason: {DropReason.COMPLICATED_LAYOUT}'
)
need_drop = True
drop_reason.append(DropReason.COMPLICATED_LAYOUT)
layout_column_width = get_columns_cnt_of_layout(layout_tree)
if layout_column_width > 2: # 去掉超过2列的布局pdf
logger.warning(
f'skip this page, page_id: {page_id}, reason: {DropReason.TOO_MANY_LAYOUT_COLUMNS}'
)
need_drop = True
drop_reason.append(DropReason.TOO_MANY_LAYOUT_COLUMNS)
"""根据layout顺序，对当前页面所有需要留下的block进行排序"""
sorted_blocks = sort_blocks_by_layout(all_bboxes, layout_bboxes)
"""将span填入排好序的blocks中"""
block_with_spans, spans = fill_spans_in_blocks(sorted_blocks, spans, 0.3)
"""对block进行fix操作"""
fix_blocks = fix_block_spans(block_with_spans, img_blocks, table_blocks)
"""获取QA需要外置的list"""
images, tables, interline_equations = get_qa_need_list_v2(fix_blocks)
"""构造pdf_info_dict"""
page_info = ocr_construct_page_component_v2(
fix_blocks,
layout_bboxes,
page_id,
page_w,
page_h,
layout_tree,
images,
tables,
interline_equations,
fix_discarded_blocks,
need_drop,
drop_reason,
)
return page_info
def pdf_parse_union(
pdf_bytes,
model_list,
imageWriter,
parse_mode,
start_page_id=0,
end_page_id=None,
debug_mode=False,
):
pdf_bytes_md5 = compute_md5(pdf_bytes)
pdf_docs = fitz.open('pdf', pdf_bytes)
"""初始化空的pdf_info_dict"""
pdf_info_dict = {}
"""用model_list和docs对象初始化magic_model"""
magic_model = MagicModel(model_list, pdf_docs)
"""根据输入的起始范围解析pdf"""
# end_page_id = end_page_id if end_page_id else len(pdf_docs) - 1
end_page_id = (
end_page_id
if end_page_id is not None and end_page_id >= 0
else len(pdf_docs) - 1
)
if end_page_id > len(pdf_docs) - 1:
logger.warning('end_page_id is out of range, use pdf_docs length')
end_page_id = len(pdf_docs) - 1
"""初始化启动时间"""
start_time = time.time()
for page_id, page in enumerate(pdf_docs):
"""debug时输出每页解析的耗时."""
if debug_mode:
time_now = time.time()
logger.info(
f'page_id: {page_id}, last_page_cost_time: {get_delta_time(start_time)}'
)
start_time = time_now
"""解析pdf中的每一页"""
if start_page_id <= page_id <= end_page_id:
page_info = parse_page_core(
pdf_docs, magic_model, page_id, pdf_bytes_md5, imageWriter, parse_mode
)
else:
page_w = page.rect.width
page_h = page.rect.height
page_info = ocr_construct_page_component_v2(
[], [], page_id, page_w, page_h, [], [], [], [], [], True, 'skip page'
)
pdf_info_dict[f'page_{page_id}'] = page_info
"""分段"""
para_split(pdf_info_dict, debug_mode=debug_mode)
"""dict转list"""
pdf_info_list = dict_to_list(pdf_info_dict)
new_pdf_info_dict = {
'pdf_info': pdf_info_list,
}
return new_pdf_info_dict
if __name__ == '__main__':
pass
magic_pdf/pdf_parse_union_core_v2.py
View file @ 8afff9ae
......@@ -5,19 +5,18 @@ import time
from typing import List
import torch
import fitz
from loguru import logger
from magic_pdf.config.drop_reason import DropReason
from magic_pdf.config.enums import SupportedPdfParseMethod
from magic_pdf.config.ocr_content_type import BlockType, ContentType
from magic_pdf.data.dataset import Dataset, PageableData
from magic_pdf.libs.boxbase import calculate_overlap_area_in_bbox1_area_ratio
from magic_pdf.libs.clean_memory import clean_memory
from magic_pdf.libs.commons import fitz, get_delta_time
from magic_pdf.libs.config_reader import get_local_layoutreader_model_dir
from magic_pdf.libs.convert_utils import dict_to_list
from magic_pdf.libs.hash_utils import compute_md5
from magic_pdf.libs.local_math import float_equal
from magic_pdf.libs.pdf_image_tools import cut_image_to_pil_image
from magic_pdf.model.magic_model import MagicModel
......@@ -34,13 +33,11 @@ except ImportError:
from magic_pdf.model.sub_modules.model_init import AtomModelSingleton
from magic_pdf.para.para_split_v3 import para_split
from magic_pdf.pre_proc.citationmarker_remove import remove_citation_marker
from magic_pdf.pre_proc.construct_page_dict import \
ocr_construct_page_component_v2
from magic_pdf.pre_proc.cut_image import ocr_cut_image_and_table
from magic_pdf.pre_proc.equations_replace import (
combine_chars_to_pymudict, remove_chars_in_text_blocks,
replace_equations_in_textblock)
from magic_pdf.pre_proc.ocr_detect_all_bboxes import \
ocr_prepare_bboxes_for_layout_split_v2
from magic_pdf.pre_proc.ocr_dict_merge import (fill_spans_in_blocks,
......@@ -49,26 +46,6 @@ from magic_pdf.pre_proc.ocr_dict_merge import (fill_spans_in_blocks,
from magic_pdf.pre_proc.ocr_span_list_modify import (
get_qa_need_list_v2, remove_overlaps_low_confidence_spans,
remove_overlaps_min_spans)
from magic_pdf.pre_proc.resolve_bbox_conflict import \
check_useful_block_horizontal_overlap
def remove_horizontal_overlap_block_which_smaller(all_bboxes):
useful_blocks = []
for bbox in all_bboxes:
useful_blocks.append({'bbox': bbox[:4]})
is_useful_block_horz_overlap, smaller_bbox, bigger_bbox = (
check_useful_block_horizontal_overlap(useful_blocks)
)
if is_useful_block_horz_overlap:
logger.warning(
f'skip this page, reason: {DropReason.USEFUL_BLOCK_HOR_OVERLAP}, smaller bbox is {smaller_bbox}, bigger bbox is {bigger_bbox}'
) # noqa: E501
for bbox in all_bboxes.copy():
if smaller_bbox == bbox[:4]:
all_bboxes.remove(bbox)
return is_useful_block_horz_overlap, all_bboxes
def __replace_STX_ETX(text_str: str):
......@@ -89,28 +66,25 @@ def __replace_STX_ETX(text_str: str):
def chars_to_content(span):
# # 先给chars按char['bbox']的x坐标排序
# span['chars'] = sorted(span['chars'], key=lambda x: x['bbox'][0])
# 检查span中的char是否为空
if len(span['chars']) == 0:
span['content'] = ''
else:
# 先给chars按char['bbox']的中心点的x坐标排序
span['chars'] = sorted(span['chars'], key=lambda x: (x['bbox'][0] + x['bbox'][2]) / 2)
content = ''
# 求char的平均宽度
if len(span['chars']) == 0:
span['content'] = content
del span['chars']
return
else:
char_width_sum = sum([char['bbox'][2] - char['bbox'][0] for char in span['chars']])
char_avg_width = char_width_sum / len(span['chars'])
content = ''
for char in span['chars']:
# 如果下一个char的x0和上一个char的x1距离超过一个字符宽度，则需要在中间插入一个空格
if char['bbox'][0] - span['chars'][span['chars'].index(char) - 1]['bbox'][2] > char_avg_width:
content += ' '
content += char['c']
span['content'] = __replace_STX_ETX(content)
del span['chars']
......@@ -128,8 +102,13 @@ def fill_char_in_spans(spans, all_chars):
span['chars'].append(char)
break
empty_spans = []
for span in spans:
chars_to_content(span)
if len(span['content']) == 0:
empty_spans.append(span)
return empty_spans
# 使用鲁棒性更强的中心点坐标判断
......@@ -162,48 +141,79 @@ def calculate_char_in_span(char_bbox, span_bbox, char_is_line_stop_flag):
def txt_spans_extract_v2(pdf_page, spans, all_bboxes, all_discarded_blocks, lang):
text_blocks_raw = pdf_page.get_text('rawdict', flags=fitz.TEXTFLAGS_TEXT)['blocks']
# @todo: 拿到char之后把倾斜角度较大的先删一遍
all_pymu_chars = []
for block in text_blocks_raw:
for line in block['lines']:
for span in line['spans']:
all_pymu_chars.extend(span['chars'])
# 计算所有sapn的高度的中位数
span_height_list = []
for span in spans:
if span['type'] in [ContentType.InterlineEquation, ContentType.Image, ContentType.Table]:
continue
span_height = span['bbox'][3] - span['bbox'][1]
span['height'] = span_height
span_height_list.append(span_height)
if len(span_height_list) == 0:
return spans
else:
median_span_height = statistics.median(span_height_list)
useful_spans = []
unuseful_spans = []
# 纵向span的两个特征：1. 高度超过多个line 2. 高宽比超过某个值
vertical_spans = []
for span in spans:
for block in all_bboxes:
if span['type'] in [ContentType.InterlineEquation, ContentType.Image, ContentType.Table]:
continue
for block in all_bboxes + all_discarded_blocks:
if block[7] in [BlockType.ImageBody, BlockType.TableBody, BlockType.InterlineEquation]:
continue
else:
if calculate_overlap_area_in_bbox1_area_ratio(span['bbox'], block[0:4]) > 0.5:
if span['height'] > median_span_height * 3 and span['height'] > (span['bbox'][2] - span['bbox'][0]) * 3:
vertical_spans.append(span)
elif block in all_bboxes:
useful_spans.append(span)
break
for block in all_discarded_blocks:
if calculate_overlap_area_in_bbox1_area_ratio(span['bbox'], block[0:4]) > 0.5:
else:
unuseful_spans.append(span)
break
text_blocks = pdf_page.get_text('rawdict', flags=fitz.TEXTFLAGS_TEXT)['blocks']
del span['height']
# @todo: 拿到char之后把倾斜角度较大的先删一遍
all_pymu_chars = []
break
"""垂直的span框直接用pymu的line进行填充"""
if len(vertical_spans) > 0:
text_blocks = pdf_page.get_text('dict', flags=fitz.TEXTFLAGS_TEXT)['blocks']
all_pymu_lines = []
for block in text_blocks:
for line in block['lines']:
for span in line['spans']:
all_pymu_chars.extend(span['chars'])
all_pymu_lines.append(line)
new_spans = []
for pymu_line in all_pymu_lines:
for span in vertical_spans:
if calculate_overlap_area_in_bbox1_area_ratio(pymu_line['bbox'], span['bbox']) > 0.5:
for pymu_span in pymu_line['spans']:
span['content'] += pymu_span['text']
break
for span in useful_spans:
if span['type'] in [ContentType.Text]:
span['chars'] = []
new_spans.append(span)
for span in vertical_spans:
if len(span['content']) == 0:
spans.remove(span)
"""水平的span框如果没有char则用ocr进行填充"""
new_spans = []
for span in unuseful_spans:
for span in useful_spans + unuseful_spans:
if span['type'] in [ContentType.Text]:
span['chars'] = []
new_spans.append(span)
fill_char_in_spans(new_spans, all_pymu_chars)
empty_spans = fill_char_in_spans(new_spans, all_pymu_chars)
empty_spans = []
for span in new_spans:
if len(span['content']) == 0:
empty_spans.append(span)
if len(empty_spans) > 0:
# 初始化ocr模型
......@@ -216,52 +226,18 @@ def txt_spans_extract_v2(pdf_page, spans, all_bboxes, all_discarded_blocks, lang
)
for span in empty_spans:
spans.remove(span)
# 对span的bbox截图
# 对span的bbox截图再ocr
span_img = cut_image_to_pil_image(span['bbox'], pdf_page, mode="cv2")
ocr_res = ocr_model.ocr(span_img, det=False)
# logger.info(f"ocr_res: {ocr_res}")
# logger.info(f"empty_span: {span}")
if ocr_res and len(ocr_res) > 0:
if len(ocr_res[0]) > 0:
ocr_text, ocr_score = ocr_res[0][0]
if ocr_score > 0.5 and len(ocr_text) > 0:
span['content'] = ocr_text
spans.append(span)
return spans
span['score'] = ocr_score
else:
spans.remove(span)
def txt_spans_extract_v1(pdf_page, inline_equations, interline_equations):
text_raw_blocks = pdf_page.get_text('dict', flags=fitz.TEXTFLAGS_TEXT)['blocks']
char_level_text_blocks = pdf_page.get_text('rawdict', flags=fitz.TEXTFLAGS_TEXT)[
'blocks'
]
text_blocks = combine_chars_to_pymudict(text_raw_blocks, char_level_text_blocks)
text_blocks = replace_equations_in_textblock(
text_blocks, inline_equations, interline_equations
)
text_blocks = remove_citation_marker(text_blocks)
text_blocks = remove_chars_in_text_blocks(text_blocks)
spans = []
for v in text_blocks:
for line in v['lines']:
for span in line['spans']:
bbox = span['bbox']
if float_equal(bbox[0], bbox[2]) or float_equal(bbox[1], bbox[3]):
continue
if span.get('type') not in (
ContentType.InlineEquation,
ContentType.InterlineEquation,
):
spans.append(
{
'bbox': list(span['bbox']),
'content': __replace_STX_ETX(span['text']),
'type': ContentType.Text,
'score': 1.0,
}
)
return spans
......@@ -682,6 +658,23 @@ def parse_page_core(
"""顺便删除大水印并保留abandon的span"""
spans = remove_outside_spans(spans, all_bboxes, all_discarded_blocks)
"""删除重叠spans中置信度较低的那些"""
spans, dropped_spans_by_confidence = remove_overlaps_low_confidence_spans(spans)
"""删除重叠spans中较小的那些"""
spans, dropped_spans_by_span_overlap = remove_overlaps_min_spans(spans)
"""根据parse_mode，构造spans，主要是文本类的字符填充"""
if parse_mode == SupportedPdfParseMethod.TXT:
"""使用新版本的混合ocr方案"""
spans = txt_spans_extract_v2(page_doc, spans, all_bboxes, all_discarded_blocks, lang)
elif parse_mode == SupportedPdfParseMethod.OCR:
pass
else:
raise Exception('parse_mode must be txt or ocr')
"""先处理不需要排版的discarded_blocks"""
discarded_block_with_spans, spans = fill_spans_in_blocks(
all_discarded_blocks, spans, 0.4
......@@ -706,26 +699,6 @@ def parse_page_core(
drop_reason,
)
"""删除重叠spans中置信度较低的那些"""
spans, dropped_spans_by_confidence = remove_overlaps_low_confidence_spans(spans)
"""删除重叠spans中较小的那些"""
spans, dropped_spans_by_span_overlap = remove_overlaps_min_spans(spans)
"""根据parse_mode，构造spans，主要是文本类的字符填充"""
if parse_mode == SupportedPdfParseMethod.TXT:
"""之前的公式替换方案"""
# pymu_spans = txt_spans_extract_v1(page_doc, inline_equations, interline_equations)
# spans = replace_text_span(pymu_spans, spans)
"""ocr 中文本类的 span 用 pymu spans 替换！"""
spans = txt_spans_extract_v2(page_doc, spans, all_bboxes, all_discarded_blocks, lang)
elif parse_mode == SupportedPdfParseMethod.OCR:
pass
else:
raise Exception('parse_mode must be txt or ocr')
"""对image和table截图"""
spans = ocr_cut_image_and_table(
spans, page_doc, page_id, pdf_bytes_md5, imageWriter
......@@ -811,7 +784,7 @@ def pdf_parse_union(
if debug_mode:
time_now = time.time()
logger.info(
f'page_id: {page_id}, last_page_cost_time: {get_delta_time(start_time)}'
f'page_id: {page_id}, last_page_cost_time: {round(time.time() - start_time, 2)}'
)
start_time = time_now
......
magic_pdf/post_proc/detect_para.py deleted 100644 → 0
View file @ 4df1eb74
import os
import sys
import json
import re
import math
import unicodedata
from collections import Counter
import numpy as np
from termcolor import cprint
from magic_pdf.libs.commons import fitz
from magic_pdf.libs.nlp_utils import NLPModels
if sys.version_info[0] >= 3:
sys.stdout.reconfigure(encoding="utf-8") # type: ignore
def open_pdf(pdf_path):
try:
pdf_document = fitz.open(pdf_path) # type: ignore
return pdf_document
except Exception as e:
print(f"无法打开PDF文件：{pdf_path}。原因是：{e}")
raise e
def print_green_on_red(text):
cprint(text, "green", "on_red", attrs=["bold"], end="\n\n")
def print_green(text):
print()
cprint(text, "green", attrs=["bold"], end="\n\n")
def print_red(text):
print()
cprint(text, "red", attrs=["bold"], end="\n\n")
def print_yellow(text):
print()
cprint(text, "yellow", attrs=["bold"], end="\n\n")
def safe_get(dict_obj, key, default):
val = dict_obj.get(key)
if val is None:
return default
else:
return val
def is_bbox_overlap(bbox1, bbox2):
"""
This function checks if bbox1 and bbox2 overlap or not
Parameters
----------
bbox1 : list
bbox1
bbox2 : list
bbox2
Returns
-------
bool
True if bbox1 and bbox2 overlap, else False
"""
x0_1, y0_1, x1_1, y1_1 = bbox1
x0_2, y0_2, x1_2, y1_2 = bbox2
if x0_1 > x1_2 or x0_2 > x1_1:
return False
if y0_1 > y1_2 or y0_2 > y1_1:
return False
return True
def is_in_bbox(bbox1, bbox2):
"""
This function checks if bbox1 is in bbox2
Parameters
----------
bbox1 : list
bbox1
bbox2 : list
bbox2
Returns
-------
bool
True if bbox1 is in bbox2, else False
"""
x0_1, y0_1, x1_1, y1_1 = bbox1
x0_2, y0_2, x1_2, y1_2 = bbox2
if x0_1 >= x0_2 and y0_1 >= y0_2 and x1_1 <= x1_2 and y1_1 <= y1_2:
return True
else:
return False
def calculate_para_bbox(lines):
"""
This function calculates the minimum bbox of the paragraph
Parameters
----------
lines : list
lines
Returns
-------
para_bbox : list
bbox of the paragraph
"""
x0 = min(line["bbox"][0] for line in lines)
y0 = min(line["bbox"][1] for line in lines)
x1 = max(line["bbox"][2] for line in lines)
y1 = max(line["bbox"][3] for line in lines)
return [x0, y0, x1, y1]
def is_line_right_aligned_from_neighbors(curr_line_bbox, prev_line_bbox, next_line_bbox, avg_char_width, direction=2):
"""
This function checks if the line is right aligned from its neighbors
Parameters
----------
curr_line_bbox : list
bbox of the current line
prev_line_bbox : list
bbox of the previous line
next_line_bbox : list
bbox of the next line
avg_char_width : float
average of char widths
direction : int
0 for prev, 1 for next, 2 for both
Returns
-------
bool
True if the line is right aligned from its neighbors, False otherwise.
"""
horizontal_ratio = 0.5
horizontal_thres = horizontal_ratio * avg_char_width
_, _, x1, _ = curr_line_bbox
_, _, prev_x1, _ = prev_line_bbox if prev_line_bbox else (0, 0, 0, 0)
_, _, next_x1, _ = next_line_bbox if next_line_bbox else (0, 0, 0, 0)
if direction == 0:
return abs(x1 - prev_x1) < horizontal_thres
elif direction == 1:
return abs(x1 - next_x1) < horizontal_thres
elif direction == 2:
return abs(x1 - prev_x1) < horizontal_thres and abs(x1 - next_x1) < horizontal_thres
else:
return False
def is_line_left_aligned_from_neighbors(curr_line_bbox, prev_line_bbox, next_line_bbox, avg_char_width, direction=2):
"""
This function checks if the line is left aligned from its neighbors
Parameters
----------
curr_line_bbox : list
bbox of the current line
prev_line_bbox : list
bbox of the previous line
next_line_bbox : list
bbox of the next line
avg_char_width : float
average of char widths
direction : int
0 for prev, 1 for next, 2 for both
Returns
-------
bool
True if the line is left aligned from its neighbors, False otherwise.
"""
horizontal_ratio = 0.5
horizontal_thres = horizontal_ratio * avg_char_width
x0, _, _, _ = curr_line_bbox
prev_x0, _, _, _ = prev_line_bbox if prev_line_bbox else (0, 0, 0, 0)
next_x0, _, _, _ = next_line_bbox if next_line_bbox else (0, 0, 0, 0)
if direction == 0:
return abs(x0 - prev_x0) < horizontal_thres
elif direction == 1:
return abs(x0 - next_x0) < horizontal_thres
elif direction == 2:
return abs(x0 - prev_x0) < horizontal_thres and abs(x0 - next_x0) < horizontal_thres
else:
return False
def end_with_punctuation(line_text):
"""
This function checks if the line ends with punctuation marks
"""
english_end_puncs = [".", "?", "!"]
chinese_end_puncs = ["。", "？", "！"]
end_puncs = english_end_puncs + chinese_end_puncs
last_non_space_char = None
for ch in line_text[::-1]:
if not ch.isspace():
last_non_space_char = ch
break
if last_non_space_char is None:
return False
return last_non_space_char in end_puncs
def is_nested_list(lst):
if isinstance(lst, list):
return any(isinstance(sub, list) for sub in lst)
return False
class DenseSingleLineBlockException(Exception):
"""
This class defines the exception type for dense single line-block.
"""
def __init__(self, message="DenseSingleLineBlockException"):
self.message = message
super().__init__(self.message)
def __str__(self):
return f"{self.message}"
def __repr__(self):
return f"{self.message}"
class TitleDetectionException(Exception):
"""
This class defines the exception type for title detection.
"""
def __init__(self, message="TitleDetectionException"):
self.message = message
super().__init__(self.message)
def __str__(self):
return f"{self.message}"
def __repr__(self):
return f"{self.message}"
class TitleLevelException(Exception):
"""
This class defines the exception type for title level.
"""
def __init__(self, message="TitleLevelException"):
self.message = message
super().__init__(self.message)
def __str__(self):
return f"{self.message}"
def __repr__(self):
return f"{self.message}"
class ParaSplitException(Exception):
"""
This class defines the exception type for paragraph splitting.
"""
def __init__(self, message="ParaSplitException"):
self.message = message
super().__init__(self.message)
def __str__(self):
return f"{self.message}"
def __repr__(self):
return f"{self.message}"
class ParaMergeException(Exception):
"""
This class defines the exception type for paragraph merging.
"""
def __init__(self, message="ParaMergeException"):
self.message = message
super().__init__(self.message)
def __str__(self):
return f"{self.message}"
def __repr__(self):
return f"{self.message}"
class DiscardByException:
"""
This class discards pdf files by exception
"""
def __init__(self) -> None:
pass
def discard_by_single_line_block(self, pdf_dic, exception: DenseSingleLineBlockException):
"""
This function discards pdf files by single line block exception
Parameters
----------
pdf_dic : dict
pdf dictionary
exception : str
exception message
Returns
-------
error_message : str
"""
exception_page_nums = 0
page_num = 0
for page_id, page in pdf_dic.items():
if page_id.startswith("page_"):
page_num += 1
if "preproc_blocks" in page.keys():
preproc_blocks = page["preproc_blocks"]
all_single_line_blocks = []
for block in preproc_blocks:
if len(block["lines"]) == 1:
all_single_line_blocks.append(block)
if len(preproc_blocks) > 0 and len(all_single_line_blocks) / len(preproc_blocks) > 0.9:
exception_page_nums += 1
if page_num == 0:
return None
if exception_page_nums / page_num > 0.1: # Low ratio means basically, whenever this is the case, it is discarded
return exception.message
return None
def discard_by_title_detection(self, pdf_dic, exception: TitleDetectionException):
"""
This function discards pdf files by title detection exception
Parameters
----------
pdf_dic : dict
pdf dictionary
exception : str
exception message
Returns
-------
error_message : str
"""
# return exception.message
return None
def discard_by_title_level(self, pdf_dic, exception: TitleLevelException):
"""
This function discards pdf files by title level exception
Parameters
----------
pdf_dic : dict
pdf dictionary
exception : str
exception message
Returns
-------
error_message : str
"""
# return exception.message
return None
def discard_by_split_para(self, pdf_dic, exception: ParaSplitException):
"""
This function discards pdf files by split para exception
Parameters
----------
pdf_dic : dict
pdf dictionary
exception : str
exception message
Returns
-------
error_message : str
"""
# return exception.message
return None
def discard_by_merge_para(self, pdf_dic, exception: ParaMergeException):
"""
This function discards pdf files by merge para exception
Parameters
----------
pdf_dic : dict
pdf dictionary
exception : str
exception message
Returns
-------
error_message : str
"""
# return exception.message
return None
class LayoutFilterProcessor:
def __init__(self) -> None:
pass
def batch_process_blocks(self, pdf_dict):
"""
This function processes the blocks in batch.
Parameters
----------
self : object
The instance of the class.
pdf_dict : dict
pdf dictionary
Returns
-------
pdf_dict : dict
pdf dictionary
"""
for page_id, blocks in pdf_dict.items():
if page_id.startswith("page_"):
if "layout_bboxes" in blocks.keys() and "para_blocks" in blocks.keys():
layout_bbox_objs = blocks["layout_bboxes"]
if layout_bbox_objs is None:
continue
layout_bboxes = [bbox_obj["layout_bbox"] for bbox_obj in layout_bbox_objs]
# Enlarge each value of x0, y0, x1, y1 for each layout_bbox to prevent loss of text.
layout_bboxes = [
[math.ceil(x0), math.ceil(y0), math.ceil(x1), math.ceil(y1)] for x0, y0, x1, y1 in layout_bboxes
]
para_blocks = blocks["para_blocks"]
if para_blocks is None:
continue
for lb_bbox in layout_bboxes:
for i, para_block in enumerate(para_blocks):
para_bbox = para_block["bbox"]
para_blocks[i]["in_layout"] = 0
if is_in_bbox(para_bbox, lb_bbox):
para_blocks[i]["in_layout"] = 1
blocks["para_blocks"] = para_blocks
return pdf_dict
class RawBlockProcessor:
def __init__(self) -> None:
self.y_tolerance = 2
self.pdf_dic = {}
def __span_flags_decomposer(self, span_flags):
"""
Make font flags human readable.
Parameters
----------
self : object
The instance of the class.
span_flags : int
span flags
Returns
-------
l : dict
decomposed flags
"""
l = {
"is_superscript": False,
"is_italic": False,
"is_serifed": False,
"is_sans_serifed": False,
"is_monospaced": False,
"is_proportional": False,
"is_bold": False,
}
if span_flags & 2**0:
l["is_superscript"] = True # 表示上标
if span_flags & 2**1:
l["is_italic"] = True # 表示斜体
if span_flags & 2**2:
l["is_serifed"] = True # 表示衬线字体
else:
l["is_sans_serifed"] = True # 表示非衬线字体
if span_flags & 2**3:
l["is_monospaced"] = True # 表示等宽字体
else:
l["is_proportional"] = True # 表示比例字体
if span_flags & 2**4:
l["is_bold"] = True # 表示粗体
return l
def __make_new_lines(self, raw_lines):
"""
This function makes new lines.
Parameters
----------
self : object
The instance of the class.
raw_lines : list
raw lines
Returns
-------
new_lines : list
new lines
"""
new_lines = []
new_line = None
for raw_line in raw_lines:
raw_line_bbox = raw_line["bbox"]
raw_line_spans = raw_line["spans"]
raw_line_text = "".join([span["text"] for span in raw_line_spans])
raw_line_dir = raw_line.get("dir", None)
decomposed_line_spans = []
for span in raw_line_spans:
raw_flags = span["flags"]
decomposed_flags = self.__span_flags_decomposer(raw_flags)
span["decomposed_flags"] = decomposed_flags
decomposed_line_spans.append(span)
if new_line is None: # Handle the first line
new_line = {
"bbox": raw_line_bbox,
"text": raw_line_text,
"dir": raw_line_dir if raw_line_dir else (0, 0),
"spans": decomposed_line_spans,
}
else: # Handle the rest lines
if (
abs(raw_line_bbox[1] - new_line["bbox"][1]) <= self.y_tolerance
and abs(raw_line_bbox[3] - new_line["bbox"][3]) <= self.y_tolerance
):
new_line["bbox"] = (
min(new_line["bbox"][0], raw_line_bbox[0]), # left
new_line["bbox"][1], # top
max(new_line["bbox"][2], raw_line_bbox[2]), # right
raw_line_bbox[3], # bottom
)
new_line["text"] += raw_line_text
new_line["spans"].extend(raw_line_spans)
new_line["dir"] = (
new_line["dir"][0] + raw_line_dir[0],
new_line["dir"][1] + raw_line_dir[1],
)
else:
new_lines.append(new_line)
new_line = {
"bbox": raw_line_bbox,
"text": raw_line_text,
"dir": raw_line_dir if raw_line_dir else (0, 0),
"spans": raw_line_spans,
}
if new_line:
new_lines.append(new_line)
return new_lines
def __make_new_block(self, raw_block):
"""
This function makes a new block.
Parameters
----------
self : object
The instance of the class.
----------
raw_block : dict
a raw block
Returns
-------
new_block : dict
"""
new_block = {}
block_id = raw_block["number"]
block_bbox = raw_block["bbox"]
block_text = "".join(span["text"] for line in raw_block["lines"] for span in line["spans"])
raw_lines = raw_block["lines"]
block_lines = self.__make_new_lines(raw_lines)
new_block["block_id"] = block_id
new_block["bbox"] = block_bbox
new_block["text"] = block_text
new_block["lines"] = block_lines
return new_block
def batch_process_blocks(self, pdf_dic):
"""
This function processes the blocks in batch.
Parameters
----------
self : object
The instance of the class.
----------
blocks : list
Input block is a list of raw blocks.
Returns
-------
result_dict : dict
result dictionary
"""
for page_id, blocks in pdf_dic.items():
if page_id.startswith("page_"):
para_blocks = []
if "preproc_blocks" in blocks.keys():
input_blocks = blocks["preproc_blocks"]
for raw_block in input_blocks:
new_block = self.__make_new_block(raw_block)
para_blocks.append(new_block)
blocks["para_blocks"] = para_blocks
return pdf_dic
class BlockStatisticsCalculator:
"""
This class calculates the statistics of the block.
"""
def __init__(self) -> None:
pass
def __calc_stats_of_new_lines(self, new_lines):
"""
This function calculates the paragraph metrics
Parameters
----------
combined_lines : list
combined lines
Returns
-------
X0 : float
Median of x0 values, which represents the left average boundary of the block
X1 : float
Median of x1 values, which represents the right average boundary of the block
avg_char_width : float
Average of char widths, which represents the average char width of the block
avg_char_height : float
Average of line heights, which represents the average line height of the block
"""
x0_values = []
x1_values = []
char_widths = []
char_heights = []
block_font_types = []
block_font_sizes = []
block_directions = []
if len(new_lines) > 0:
for i, line in enumerate(new_lines):
line_bbox = line["bbox"]
line_text = line["text"]
line_spans = line["spans"]
num_chars = len([ch for ch in line_text if not ch.isspace()])
x0_values.append(line_bbox[0])
x1_values.append(line_bbox[2])
if num_chars > 0:
char_width = (line_bbox[2] - line_bbox[0]) / num_chars
char_widths.append(char_width)
for span in line_spans:
block_font_types.append(span["font"])
block_font_sizes.append(span["size"])
if "dir" in line:
block_directions.append(line["dir"])
# line_font_types = [span["font"] for span in line_spans]
char_heights = [span["size"] for span in line_spans]
X0 = np.median(x0_values) if x0_values else 0
X1 = np.median(x1_values) if x1_values else 0
avg_char_width = sum(char_widths) / len(char_widths) if char_widths else 0
avg_char_height = sum(char_heights) / len(char_heights) if char_heights else 0
# max_freq_font_type = max(set(block_font_types), key=block_font_types.count) if block_font_types else None
max_span_length = 0
max_span_font_type = None
for line in new_lines:
line_spans = line["spans"]
for span in line_spans:
span_length = span["bbox"][2] - span["bbox"][0]
if span_length > max_span_length:
max_span_length = span_length
max_span_font_type = span["font"]
max_freq_font_type = max_span_font_type
avg_font_size = sum(block_font_sizes) / len(block_font_sizes) if block_font_sizes else None
avg_dir_horizontal = sum([dir[0] for dir in block_directions]) / len(block_directions) if block_directions else 0
avg_dir_vertical = sum([dir[1] for dir in block_directions]) / len(block_directions) if block_directions else 0
median_font_size = float(np.median(block_font_sizes)) if block_font_sizes else None
return (
X0,
X1,
avg_char_width,
avg_char_height,
max_freq_font_type,
avg_font_size,
(avg_dir_horizontal, avg_dir_vertical),
median_font_size,
)
def __make_new_block(self, input_block):
new_block = {}
raw_lines = input_block["lines"]
stats = self.__calc_stats_of_new_lines(raw_lines)
block_id = input_block["block_id"]
block_bbox = input_block["bbox"]
block_text = input_block["text"]
block_lines = raw_lines
block_avg_left_boundary = stats[0]
block_avg_right_boundary = stats[1]
block_avg_char_width = stats[2]
block_avg_char_height = stats[3]
block_font_type = stats[4]
block_font_size = stats[5]
block_direction = stats[6]
block_median_font_size = stats[7]
new_block["block_id"] = block_id
new_block["bbox"] = block_bbox
new_block["text"] = block_text
new_block["dir"] = block_direction
new_block["X0"] = block_avg_left_boundary
new_block["X1"] = block_avg_right_boundary
new_block["avg_char_width"] = block_avg_char_width
new_block["avg_char_height"] = block_avg_char_height
new_block["block_font_type"] = block_font_type
new_block["block_font_size"] = block_font_size
new_block["lines"] = block_lines
new_block["median_font_size"] = block_median_font_size
return new_block
def batch_process_blocks(self, pdf_dic):
"""
This function processes the blocks in batch.
Parameters
----------
self : object
The instance of the class.
----------
blocks : list
Input block is a list of raw blocks.
Schema can refer to the value of key ""preproc_blocks".
Returns
-------
result_dict : dict
result dictionary
"""
for page_id, blocks in pdf_dic.items():
if page_id.startswith("page_"):
para_blocks = []
if "para_blocks" in blocks.keys():
input_blocks = blocks["para_blocks"]
for input_block in input_blocks:
new_block = self.__make_new_block(input_block)
para_blocks.append(new_block)
blocks["para_blocks"] = para_blocks
return pdf_dic
class DocStatisticsCalculator:
"""
This class calculates the statistics of the document.
"""
def __init__(self) -> None:
pass
def calc_stats_of_doc(self, pdf_dict):
"""
This function computes the statistics of the document
Parameters
----------
result_dict : dict
result dictionary
Returns
-------
statistics : dict
statistics of the document
"""
total_text_length = 0
total_num_blocks = 0
for page_id, blocks in pdf_dict.items():
if page_id.startswith("page_"):
if "para_blocks" in blocks.keys():
para_blocks = blocks["para_blocks"]
for para_block in para_blocks:
total_text_length += len(para_block["text"])
total_num_blocks += 1
avg_text_length = total_text_length / total_num_blocks if total_num_blocks else 0
font_list = []
for page_id, blocks in pdf_dict.items():
if page_id.startswith("page_"):
if "para_blocks" in blocks.keys():
input_blocks = blocks["para_blocks"]
for input_block in input_blocks:
block_text_length = len(input_block.get("text", ""))
if block_text_length < avg_text_length * 0.5:
continue
block_font_type = safe_get(input_block, "block_font_type", "")
block_font_size = safe_get(input_block, "block_font_size", 0)
font_list.append((block_font_type, block_font_size))
font_counter = Counter(font_list)
most_common_font = font_counter.most_common(1)[0] if font_list else (("", 0), 0)
second_most_common_font = font_counter.most_common(2)[1] if len(font_counter) > 1 else (("", 0), 0)
statistics = {
"num_pages": 0,
"num_blocks": 0,
"num_paras": 0,
"num_titles": 0,
"num_header_blocks": 0,
"num_footer_blocks": 0,
"num_watermark_blocks": 0,
"num_vertical_margin_note_blocks": 0,
"most_common_font_type": most_common_font[0][0],
"most_common_font_size": most_common_font[0][1],
"number_of_most_common_font": most_common_font[1],
"second_most_common_font_type": second_most_common_font[0][0],
"second_most_common_font_size": second_most_common_font[0][1],
"number_of_second_most_common_font": second_most_common_font[1],
"avg_text_length": avg_text_length,
}
for page_id, blocks in pdf_dict.items():
if page_id.startswith("page_"):
blocks = pdf_dict[page_id]["para_blocks"]
statistics["num_pages"] += 1
for block_id, block_data in enumerate(blocks):
statistics["num_blocks"] += 1
if "paras" in block_data.keys():
statistics["num_paras"] += len(block_data["paras"])
for line in block_data["lines"]:
if line.get("is_title", 0):
statistics["num_titles"] += 1
if block_data.get("is_header", 0):
statistics["num_header_blocks"] += 1
if block_data.get("is_footer", 0):
statistics["num_footer_blocks"] += 1
if block_data.get("is_watermark", 0):
statistics["num_watermark_blocks"] += 1
if block_data.get("is_vertical_margin_note", 0):
statistics["num_vertical_margin_note_blocks"] += 1
pdf_dict["statistics"] = statistics
return pdf_dict
class TitleProcessor:
"""
This class processes the title.
"""
def __init__(self, *doc_statistics) -> None:
if len(doc_statistics) > 0:
self.doc_statistics = doc_statistics[0]
self.nlp_model = NLPModels()
self.MAX_TITLE_LEVEL = 3
self.numbered_title_pattern = r"""
^ # 行首
( # 开始捕获组
[\(\（]\d+[\)\）] # 括号内数字，支持中文和英文括号，例如：(1) 或 （1）
|\d+[\)\）]\s # 数字后跟右括号和空格，支持中文和英文括号，例如：2) 或 2）
|[\(\（][A-Z][\)\）] # 括号内大写字母，支持中文和英文括号，例如：(A) 或 （A）
|[A-Z][\)\）]\s # 大写字母后跟右括号和空格，例如：A) 或 A）
|[\(\（][IVXLCDM]+[\)\）] # 括号内罗马数字，支持中文和英文括号，例如：(I) 或 （I）
|[IVXLCDM]+[\)\）]\s # 罗马数字后跟右括号和空格，例如：I) 或 I）
|\d+(\.\d+)*\s # 数字或复合数字编号后跟空格，例如：1. 或 3.2.1
|[一二三四五六七八九十百千]+[、\s] # 中文序号后跟顿号和空格，例如：一、
|[\（|\(][一二三四五六七八九十百千]+[\）|\)]\s* # 中文括号内中文序号后跟空格，例如：（一）
|[A-Z]\.\d+(\.\d+)?\s # 大写字母后跟点和数字，例如：A.1 或 A.1.1
|[\(\（][a-z][\)\）] # 括号内小写字母，支持中文和英文括号，例如：(a) 或 （a）
|[a-z]\)\s # 小写字母后跟右括号和空格，例如：a)
|[A-Z]-\s # 大写字母后跟短横线和空格，例如：A-
|\w+:\s # 英文序号词后跟冒号和空格，例如：First:
|第[一二三四五六七八九十百千]+[章节部分条款]\s # 以“第”开头的中文标题后跟空格
|[IVXLCDM]+\. # 罗马数字后跟点，例如：I.
|\d+\.\s # 单个数字后跟点和空格，例如：1.
) # 结束捕获组
.+ # 标题的其余部分
"""
def _is_potential_title(
self,
curr_line,
prev_line,
prev_line_is_title,
next_line,
avg_char_width,
avg_char_height,
median_font_size,
):
"""
This function checks if the line is a potential title.
Parameters
----------
curr_line : dict
current line
prev_line : dict
previous line
next_line : dict
next line
avg_char_width : float
average of char widths
avg_char_height : float
average of line heights
Returns
-------
bool
True if the line is a potential title, False otherwise.
"""
def __is_line_centered(line_bbox, page_bbox, avg_char_width):
"""
This function checks if the line is centered on the page
Parameters
----------
line_bbox : list
bbox of the line
page_bbox : list
bbox of the page
avg_char_width : float
average of char widths
Returns
-------
bool
True if the line is centered on the page, False otherwise.
"""
horizontal_ratio = 0.5
horizontal_thres = horizontal_ratio * avg_char_width
x0, _, x1, _ = line_bbox
_, _, page_x1, _ = page_bbox
return abs((x0 + x1) / 2 - page_x1 / 2) < horizontal_thres
def __is_bold_font_line(line):
"""
Check if a line contains any bold font style.
"""
def _is_bold_span(span):
# if span text is empty or only contains space, return False
if not span["text"].strip():
return False
return bool(span["flags"] & 2**4) # Check if the font is bold
for span in line["spans"]:
if not _is_bold_span(span):
return False
return True
def __is_italic_font_line(line):
"""
Check if a line contains any italic font style.
"""
def __is_italic_span(span):
return bool(span["flags"] & 2**1) # Check if the font is italic
for span in line["spans"]:
if not __is_italic_span(span):
return False
return True
def __is_punctuation_heavy(line_text):
"""
Check if the line contains a high ratio of punctuation marks, which may indicate
that the line is not a title.
Parameters:
line_text (str): Text of the line.
Returns:
bool: True if the line is heavy with punctuation, False otherwise.
"""
# Pattern for common title format like "X.Y. Title"
pattern = r"\b\d+\.\d+\..*\b"
# If the line matches the title format, return False
if re.match(pattern, line_text.strip()):
return False
# Find all punctuation marks in the line
punctuation_marks = re.findall(r"[^\w\s]", line_text)
number_of_punctuation_marks = len(punctuation_marks)
text_length = len(line_text)
if text_length == 0:
return False
punctuation_ratio = number_of_punctuation_marks / text_length
if punctuation_ratio >= 0.1:
return True
return False
def __has_mixed_font_styles(spans, strict_mode=False):
"""
This function checks if the line has mixed font styles, the strict mode will compare the font types
Parameters
----------
spans : list
spans of the line
strict_mode : bool
True for strict mode, the font types will be fully compared
False for non-strict mode, the font types will be compared by the most longest common prefix
Returns
-------
bool
True if the line has mixed font styles, False otherwise.
"""
if strict_mode:
font_styles = set()
for span in spans:
font_style = span["font"].lower()
font_styles.add(font_style)
return len(font_styles) > 1
else: # non-strict mode
font_styles = []
for span in spans:
font_style = span["font"].lower()
font_styles.append(font_style)
if len(font_styles) > 1:
longest_common_prefix = os.path.commonprefix(font_styles)
if len(longest_common_prefix) > 0:
return False
else:
return True
else:
return False
def __is_different_font_type_from_neighbors(curr_line_font_type, prev_line_font_type, next_line_font_type):
"""
This function checks if the current line has a different font type from the previous and next lines
Parameters
----------
curr_line_font_type : str
font type of the current line
prev_line_font_type : str
font type of the previous line
next_line_font_type : str
font type of the next line
Returns
-------
bool
True if the current line has a different font type from the previous and next lines, False otherwise.
"""
return all(
curr_line_font_type != other_font_type.lower()
for other_font_type in [prev_line_font_type, next_line_font_type]
if other_font_type is not None
)
def __is_larger_font_size_from_neighbors(curr_line_font_size, prev_line_font_size, next_line_font_size):
"""
This function checks if the current line has a larger font size than the previous and next lines
Parameters
----------
curr_line_font_size : float
font size of the current line
prev_line_font_size : float
font size of the previous line
next_line_font_size : float
font size of the next line
Returns
-------
bool
True if the current line has a larger font size than the previous and next lines, False otherwise.
"""
return all(
curr_line_font_size > other_font_size * 1.2
for other_font_size in [prev_line_font_size, next_line_font_size]
if other_font_size is not None
)
def __is_similar_to_pre_line(curr_line_font_type, prev_line_font_type, curr_line_font_size, prev_line_font_size):
"""
This function checks if the current line is similar to the previous line
Parameters
----------
curr_line : dict
current line
prev_line : dict
previous line
Returns
-------
bool
True if the current line is similar to the previous line, False otherwise.
"""
if curr_line_font_type == prev_line_font_type and curr_line_font_size == prev_line_font_size:
return True
else:
return False
def __is_same_font_type_of_docAvg(curr_line_font_type):
"""
This function checks if the current line has the same font type as the document average font type
Parameters
----------
curr_line_font_type : str
font type of the current line
Returns
-------
bool
True if the current line has the same font type as the document average font type, False otherwise.
"""
doc_most_common_font_type = safe_get(self.doc_statistics, "most_common_font_type", "").lower()
doc_second_most_common_font_type = safe_get(self.doc_statistics, "second_most_common_font_type", "").lower()
return curr_line_font_type.lower() in [doc_most_common_font_type, doc_second_most_common_font_type]
def __is_font_size_not_less_than_docAvg(curr_line_font_size, ratio: float = 1):
"""
This function checks if the current line has a large enough font size
Parameters
----------
curr_line_font_size : float
font size of the current line
ratio : float
ratio of the current line font size to the document average font size
Returns
-------
bool
True if the current line has a large enough font size, False otherwise.
"""
doc_most_common_font_size = safe_get(self.doc_statistics, "most_common_font_size", 0)
doc_second_most_common_font_size = safe_get(self.doc_statistics, "second_most_common_font_size", 0)
doc_avg_font_size = min(doc_most_common_font_size, doc_second_most_common_font_size)
return curr_line_font_size >= doc_avg_font_size * ratio
def __is_sufficient_spacing_above_and_below(
curr_line_bbox,
prev_line_bbox,
next_line_bbox,
avg_char_height,
median_font_size,
):
"""
This function checks if the current line has sufficient spacing above and below
Parameters
----------
curr_line_bbox : list
bbox of the current line
prev_line_bbox : list
bbox of the previous line
next_line_bbox : list
bbox of the next line
avg_char_width : float
average of char widths
avg_char_height : float
average of line heights
Returns
-------
bool
True if the current line has sufficient spacing above and below, False otherwise.
"""
vertical_ratio = 1.25
vertical_thres = vertical_ratio * median_font_size
_, y0, _, y1 = curr_line_bbox
sufficient_spacing_above = False
if prev_line_bbox:
vertical_spacing_above = min(y0 - prev_line_bbox[1], y1 - prev_line_bbox[3])
sufficient_spacing_above = vertical_spacing_above > vertical_thres
else:
sufficient_spacing_above = True
sufficient_spacing_below = False
if next_line_bbox:
vertical_spacing_below = min(next_line_bbox[1] - y0, next_line_bbox[3] - y1)
sufficient_spacing_below = vertical_spacing_below > vertical_thres
else:
sufficient_spacing_below = True
return (sufficient_spacing_above, sufficient_spacing_below)
def __is_word_list_line_by_rules(curr_line_text):
"""
This function checks if the current line is a word list
Parameters
----------
curr_line_text : str
text of the current line
Returns
-------
bool
True if the current line is a name list, False otherwise.
"""
# name_list_pattern = r"([a-zA-Z][a-zA-Z\s]{0,20}[a-zA-Z]|[\u4e00-\u9fa5·]{2,16})(?=[，,;；\s]|$)"
name_list_pattern = r"(?<![\u4e00-\u9fa5])([A-Z][a-z]{0,19}\s[A-Z][a-z]{0,19}|[\u4e00-\u9fa5]{2,6})(?=[，,;；\s]|$)"
compiled_pattern = re.compile(name_list_pattern)
if compiled_pattern.search(curr_line_text):
return True
else:
return False
def __get_text_catgr_by_nlp(curr_line_text):
"""
This function checks if the current line is a name list using nlp model, such as spacy
Parameters
----------
curr_line_text : str
text of the current line
Returns
-------
bool
True if the current line is a name list, False otherwise.
"""
result = self.nlp_model.detect_entity_catgr_using_nlp(curr_line_text)
return result
def __is_numbered_title(curr_line_text):
"""
This function checks if the current line is a numbered list
Parameters
----------
curr_line_text : str
text of the current line
Returns
-------
bool
True if the current line is a numbered list, False otherwise.
"""
compiled_pattern = re.compile(self.numbered_title_pattern, re.VERBOSE)
if compiled_pattern.search(curr_line_text):
return True
else:
return False
def __is_end_with_ending_puncs(line_text):
"""
This function checks if the current line ends with a ending punctuation mark
Parameters
----------
line_text : str
text of the current line
Returns
-------
bool
True if the current line ends with a punctuation mark, False otherwise.
"""
end_puncs = [".", "?", "!", "。", "？", "！", "…"]
line_text = line_text.rstrip()
if line_text[-1] in end_puncs:
return True
return False
def __contains_only_no_meaning_symbols(line_text):
"""
This function checks if the current line contains only symbols that have no meaning, if so, it is not a title.
Situation contains:
1. Only have punctuation marks
2. Only have other non-meaning symbols
Parameters
----------
line_text : str
text of the current line
Returns
-------
bool
True if the current line contains only symbols that have no meaning, False otherwise.
"""
punctuation_marks = re.findall(r"[^\w\s]", line_text) # find all punctuation marks
number_of_punctuation_marks = len(punctuation_marks)
text_length = len(line_text)
if text_length == 0:
return False
punctuation_ratio = number_of_punctuation_marks / text_length
if punctuation_ratio >= 0.9:
return True
return False
def __is_equation(line_text):
"""
This function checks if the current line is an equation.
Parameters
----------
line_text : str
Returns
-------
bool
True if the current line is an equation, False otherwise.
"""
equation_reg = r"\$.*?\\overline.*?\$" # to match interline equations
if re.search(equation_reg, line_text):
return True
else:
return False
def __is_title_by_len(text, max_length=200):
"""
This function checks if the current line is a title by length.
Parameters
----------
text : str
text of the current line
max_length : int
max length of the title
Returns
-------
bool
True if the current line is a title, False otherwise.
"""
text = text.strip()
return len(text) <= max_length
def __compute_line_font_type_and_size(curr_line):
"""
This function computes the font type and font size of the line.
Parameters
----------
line : dict
line
Returns
-------
font_type : str
font type of the line
font_size : float
font size of the line
"""
spans = curr_line["spans"]
max_accumulated_length = 0
max_span_font_size = curr_line["spans"][0]["size"] # default value, float type
max_span_font_type = curr_line["spans"][0]["font"].lower() # default value, string type
for span in spans:
if span["text"].isspace():
continue
span_length = span["bbox"][2] - span["bbox"][0]
if span_length > max_accumulated_length:
max_accumulated_length = span_length
max_span_font_size = span["size"]
max_span_font_type = span["font"].lower()
return max_span_font_type, max_span_font_size
def __is_a_consistent_sub_title(pre_line, curr_line):
"""
This function checks if the current line is a consistent sub title.
Parameters
----------
pre_line : dict
previous line
curr_line : dict
current line
Returns
-------
bool
True if the current line is a consistent sub title, False otherwise.
"""
if pre_line is None:
return False
start_letter_of_pre_line = pre_line["text"][0]
start_letter_of_curr_line = curr_line["text"][0]
has_same_prefix_digit = (
start_letter_of_pre_line.isdigit()
and start_letter_of_curr_line.isdigit()
and start_letter_of_pre_line == start_letter_of_curr_line
)
# prefix text of curr_line satisfies the following title format: x.x
prefix_text_pattern = r"^\d+\.\d+"
has_subtitle_format = re.match(prefix_text_pattern, curr_line["text"])
if has_same_prefix_digit or has_subtitle_format:
return True
"""
Title detecting main Process.
"""
"""
Basic features about the current line.
"""
curr_line_bbox = curr_line["bbox"]
curr_line_text = curr_line["text"]
curr_line_font_type, curr_line_font_size = __compute_line_font_type_and_size(curr_line)
if len(curr_line_text.strip()) == 0: # skip empty lines
return False, False
prev_line_bbox = prev_line["bbox"] if prev_line else None
if prev_line:
prev_line_font_type, prev_line_font_size = __compute_line_font_type_and_size(prev_line)
else:
prev_line_font_type, prev_line_font_size = None, None
next_line_bbox = next_line["bbox"] if next_line else None
if next_line:
next_line_font_type, next_line_font_size = __compute_line_font_type_and_size(next_line)
else:
next_line_font_type, next_line_font_size = None, None
"""
Aggregated features about the current line.
"""
is_italc_font = __is_italic_font_line(curr_line)
is_bold_font = __is_bold_font_line(curr_line)
is_font_size_little_less_than_doc_avg = __is_font_size_not_less_than_docAvg(curr_line_font_size, ratio=0.8)
is_font_size_not_less_than_doc_avg = __is_font_size_not_less_than_docAvg(curr_line_font_size, ratio=1)
is_much_larger_font_than_doc_avg = __is_font_size_not_less_than_docAvg(curr_line_font_size, ratio=1.6)
is_not_same_font_type_of_docAvg = not __is_same_font_type_of_docAvg(curr_line_font_type)
is_potential_title_font = is_bold_font or is_font_size_not_less_than_doc_avg or is_not_same_font_type_of_docAvg
is_mix_font_styles_strict = __has_mixed_font_styles(curr_line["spans"], strict_mode=True)
is_mix_font_styles_loose = __has_mixed_font_styles(curr_line["spans"], strict_mode=False)
is_punctuation_heavy = __is_punctuation_heavy(curr_line_text)
is_word_list_line_by_rules = __is_word_list_line_by_rules(curr_line_text)
is_person_or_org_list_line_by_nlp = __get_text_catgr_by_nlp(curr_line_text) in ["PERSON", "GPE", "ORG"]
is_font_size_larger_than_neighbors = __is_larger_font_size_from_neighbors(
curr_line_font_size, prev_line_font_size, next_line_font_size
)
is_font_type_diff_from_neighbors = __is_different_font_type_from_neighbors(
curr_line_font_type, prev_line_font_type, next_line_font_type
)
has_sufficient_spaces_above, has_sufficient_spaces_below = __is_sufficient_spacing_above_and_below(
curr_line_bbox, prev_line_bbox, next_line_bbox, avg_char_height, median_font_size
)
is_similar_to_pre_line = __is_similar_to_pre_line(
curr_line_font_type, prev_line_font_type, curr_line_font_size, prev_line_font_size
)
is_consis_sub_title = __is_a_consistent_sub_title(prev_line, curr_line)
"""
Further aggregated features about the current line.
Attention:
Features that start with __ are for internal use.
"""
__is_line_left_aligned_from_neighbors = is_line_left_aligned_from_neighbors(
curr_line_bbox, prev_line_bbox, next_line_bbox, avg_char_width
)
__is_font_diff_from_neighbors = is_font_size_larger_than_neighbors or is_font_type_diff_from_neighbors
is_a_left_inline_title = (
is_mix_font_styles_strict and __is_line_left_aligned_from_neighbors and __is_font_diff_from_neighbors
)
is_title_by_check_prev_line = prev_line is None and has_sufficient_spaces_above and is_potential_title_font
is_title_by_check_next_line = next_line is None and has_sufficient_spaces_below and is_potential_title_font
is_title_by_check_pre_and_next_line = (
(prev_line is not None or next_line is not None)
and has_sufficient_spaces_above
and has_sufficient_spaces_below
and is_potential_title_font
)
is_numbered_title = __is_numbered_title(curr_line_text) and (
(has_sufficient_spaces_above or prev_line is None) and (has_sufficient_spaces_below or next_line is None)
)
is_not_end_with_ending_puncs = not __is_end_with_ending_puncs(curr_line_text)
is_not_only_no_meaning_symbols = not __contains_only_no_meaning_symbols(curr_line_text)
is_equation = __is_equation(curr_line_text)
is_title_by_len = __is_title_by_len(curr_line_text)
"""
Decide if the line is a title.
"""
is_title = (
is_not_end_with_ending_puncs # not end with ending punctuation marks
and is_not_only_no_meaning_symbols # not only have no meaning symbols
and is_title_by_len # is a title by length, default max length is 200
and not is_equation # an interline equation should never be a title
and is_potential_title_font # is a potential title font, which is bold or larger than the document average font size or not the same font type as the document average font type
and (
(is_not_same_font_type_of_docAvg and is_font_size_not_less_than_doc_avg)
or (is_bold_font and is_much_larger_font_than_doc_avg and is_not_same_font_type_of_docAvg)
or (
is_much_larger_font_than_doc_avg
and (is_title_by_check_prev_line or is_title_by_check_next_line or is_title_by_check_pre_and_next_line)
)
or (
is_font_size_little_less_than_doc_avg
and is_bold_font
and (is_title_by_check_prev_line or is_title_by_check_next_line or is_title_by_check_pre_and_next_line)
)
) # Consider the following situations: bold font, much larger font than doc avg, not same font type as doc avg, sufficient spacing above and below
and (
(
not is_person_or_org_list_line_by_nlp
and (
is_much_larger_font_than_doc_avg
or (is_not_same_font_type_of_docAvg and is_font_size_not_less_than_doc_avg)
)
)
or (
not (is_word_list_line_by_rules and is_person_or_org_list_line_by_nlp)
and not is_a_left_inline_title
and not is_punctuation_heavy
and (is_title_by_check_prev_line or is_title_by_check_next_line or is_title_by_check_pre_and_next_line)
)
or (
is_person_or_org_list_line_by_nlp
and (is_bold_font and is_much_larger_font_than_doc_avg and is_not_same_font_type_of_docAvg)
and (is_bold_font and is_much_larger_font_than_doc_avg and is_not_same_font_type_of_docAvg)
)
or (is_numbered_title and not is_a_left_inline_title)
) # Exclude the following situations: person/org list
)
# ) or (prev_line_is_title and is_consis_sub_title)
is_name_or_org_list_to_be_removed = (
(is_person_or_org_list_line_by_nlp)
and is_punctuation_heavy
and (is_title_by_check_prev_line or is_title_by_check_next_line or is_title_by_check_pre_and_next_line)
) and not is_title
if is_name_or_org_list_to_be_removed:
is_author_or_org_list = True
else:
is_author_or_org_list = False
# return is_title, is_author_or_org_list
"""
# print reason why the line is a title
if is_title:
print_green("This line is a title.")
print_green("↓" * 10)
print()
print("curr_line_text: ", curr_line_text)
print()
# print reason why the line is not a title
line_text = curr_line_text.strip()
test_text = "Career/Personal Life"
text_content_condition = line_text == test_text
if not is_title and text_content_condition: # Print specific line
# if not is_title: # Print each line
print_red("This line is not a title.")
print_red("↓" * 10)
print()
print("curr_line_text: ", curr_line_text)
print()
if is_not_end_with_ending_puncs:
print_green(f"is_not_end_with_ending_puncs")
else:
print_red(f"is_end_with_ending_puncs")
if is_not_only_no_meaning_symbols:
print_green(f"is_not_only_no_meaning_symbols")
else:
print_red(f"is_only_no_meaning_symbols")
if is_title_by_len:
print_green(f"is_title_by_len: {is_title_by_len}")
else:
print_red(f"is_not_title_by_len: {is_title_by_len}")
if is_equation:
print_red(f"is_equation")
else:
print_green(f"is_not_equation")
if is_potential_title_font:
print_green(f"is_potential_title_font")
else:
print_red(f"is_not_potential_title_font")
if is_punctuation_heavy:
print_red("is_punctuation_heavy")
else:
print_green("is_not_punctuation_heavy")
if is_bold_font:
print_green(f"is_bold_font")
else:
print_red(f"is_not_bold_font")
if is_font_size_not_less_than_doc_avg:
print_green(f"is_larger_font_than_doc_avg")
else:
print_red(f"is_not_larger_font_than_doc_avg")
if is_much_larger_font_than_doc_avg:
print_green(f"is_much_larger_font_than_doc_avg")
else:
print_red(f"is_not_much_larger_font_than_doc_avg")
if is_not_same_font_type_of_docAvg:
print_green(f"is_not_same_font_type_of_docAvg")
else:
print_red(f"is_same_font_type_of_docAvg")
if is_word_list_line_by_rules:
print_red("is_word_list_line_by_rules")
else:
print_green("is_not_name_list_by_rules")
if is_person_or_org_list_line_by_nlp:
print_red("is_person_or_org_list_line_by_nlp")
else:
print_green("is_not_person_or_org_list_line_by_nlp")
if not is_numbered_title:
print_red("is_not_numbered_title")
else:
print_green("is_numbered_title")
if is_a_left_inline_title:
print_red("is_a_left_inline_title")
else:
print_green("is_not_a_left_inline_title")
if not is_title_by_check_prev_line:
print_red("is_not_title_by_check_prev_line")
else:
print_green("is_title_by_check_prev_line")
if not is_title_by_check_next_line:
print_red("is_not_title_by_check_next_line")
else:
print_green("is_title_by_check_next_line")
if not is_title_by_check_pre_and_next_line:
print_red("is_not_title_by_check_pre_and_next_line")
else:
print_green("is_title_by_check_pre_and_next_line")
# print_green("Common features:")
# print_green("↓" * 10)
# print(f" curr_line_font_type: {curr_line_font_type}")
# print(f" curr_line_font_size: {curr_line_font_size}")
# print()
"""
return is_title, is_author_or_org_list
def _detect_title(self, input_block):
"""
Use the functions 'is_potential_title' to detect titles of each paragraph block.
If a line is a title, then the value of key 'is_title' of the line will be set to True.
"""
raw_lines = input_block["lines"]
prev_line_is_title_flag = False
for i, curr_line in enumerate(raw_lines):
prev_line = raw_lines[i - 1] if i > 0 else None
next_line = raw_lines[i + 1] if i < len(raw_lines) - 1 else None
blk_avg_char_width = input_block["avg_char_width"]
blk_avg_char_height = input_block["avg_char_height"]
blk_media_font_size = input_block["median_font_size"]
is_title, is_author_or_org_list = self._is_potential_title(
curr_line,
prev_line,
prev_line_is_title_flag,
next_line,
blk_avg_char_width,
blk_avg_char_height,
blk_media_font_size,
)
if is_title:
curr_line["is_title"] = is_title
prev_line_is_title_flag = True
else:
curr_line["is_title"] = False
prev_line_is_title_flag = False
# print(f"curr_line['text']: {curr_line['text']}")
# print(f"curr_line['is_title']: {curr_line['is_title']}")
# print(f"prev_line['text']: {prev_line['text'] if prev_line else None}")
# print(f"prev_line_is_title_flag: {prev_line_is_title_flag}")
# print()
if is_author_or_org_list:
curr_line["is_author_or_org_list"] = is_author_or_org_list
else:
curr_line["is_author_or_org_list"] = False
return input_block
def batch_detect_titles(self, pdf_dic):
"""
This function batch process the blocks to detect titles.
Parameters
----------
pdf_dict : dict
result dictionary
Returns
-------
pdf_dict : dict
result dictionary
"""
num_titles = 0
for page_id, blocks in pdf_dic.items():
if page_id.startswith("page_"):
para_blocks = []
if "para_blocks" in blocks.keys():
para_blocks = blocks["para_blocks"]
all_single_line_blocks = []
for block in para_blocks:
if len(block["lines"]) == 1:
all_single_line_blocks.append(block)
new_para_blocks = []
if not len(all_single_line_blocks) == len(para_blocks): # Not all blocks are single line blocks.
for para_block in para_blocks:
new_block = self._detect_title(para_block)
new_para_blocks.append(new_block)
num_titles += sum([line.get("is_title", 0) for line in new_block["lines"]])
else: # All blocks are single line blocks.
for para_block in para_blocks:
new_para_blocks.append(para_block)
num_titles += sum([line.get("is_title", 0) for line in para_block["lines"]])
para_blocks = new_para_blocks
blocks["para_blocks"] = para_blocks
for para_block in para_blocks:
all_titles = all(safe_get(line, "is_title", False) for line in para_block["lines"])
para_text_len = sum([len(line["text"]) for line in para_block["lines"]])
if (
all_titles and para_text_len < 200
): # total length of the paragraph is less than 200, more than this should not be a title
para_block["is_block_title"] = 1
else:
para_block["is_block_title"] = 0
all_name_or_org_list_to_be_removed = all(
safe_get(line, "is_author_or_org_list", False) for line in para_block["lines"]
)
if all_name_or_org_list_to_be_removed and page_id == "page_0":
para_block["is_block_an_author_or_org_list"] = 1
else:
para_block["is_block_an_author_or_org_list"] = 0
pdf_dic["statistics"]["num_titles"] = num_titles
return pdf_dic
def _recog_title_level(self, title_blocks):
"""
This function determines the title level based on the font size of the title.
Parameters
----------
title_blocks : list
Returns
-------
title_blocks : list
"""
font_sizes = np.array([safe_get(tb["block"], "block_font_size", 0) for tb in title_blocks])
# Use the mean and std of font sizes to remove extreme values
mean_font_size = np.mean(font_sizes)
std_font_size = np.std(font_sizes)
min_extreme_font_size = mean_font_size - std_font_size # type: ignore
max_extreme_font_size = mean_font_size + std_font_size # type: ignore
# Compute the threshold for title level
middle_font_sizes = font_sizes[(font_sizes > min_extreme_font_size) & (font_sizes < max_extreme_font_size)]
if middle_font_sizes.size > 0:
middle_mean_font_size = np.mean(middle_font_sizes)
level_threshold = middle_mean_font_size
else:
level_threshold = mean_font_size
for tb in title_blocks:
title_block = tb["block"]
title_font_size = safe_get(title_block, "block_font_size", 0)
current_level = 1 # Initialize title level, the biggest level is 1
# print(f"Before adjustment by font size, {current_level}")
if title_font_size >= max_extreme_font_size:
current_level = 1
elif title_font_size <= min_extreme_font_size:
current_level = 3
elif float(title_font_size) >= float(level_threshold):
current_level = 2
else:
current_level = 3
# print(f"After adjustment by font size, {current_level}")
title_block["block_title_level"] = current_level
return title_blocks
def batch_recog_title_level(self, pdf_dic):
"""
This function batch process the blocks to recognize title level.
Parameters
----------
pdf_dict : dict
result dictionary
Returns
-------
pdf_dict : dict
result dictionary
"""
title_blocks = []
# Collect all titles
for page_id, blocks in pdf_dic.items():
if page_id.startswith("page_"):
para_blocks = blocks.get("para_blocks", [])
for block in para_blocks:
if block.get("is_block_title"):
title_obj = {"page_id": page_id, "block": block}
title_blocks.append(title_obj)
# Determine title level
if title_blocks:
# Determine title level based on font size
title_blocks = self._recog_title_level(title_blocks)
return pdf_dic
class BlockTerminationProcessor:
"""
This class is used to process the block termination.
"""
def __init__(self) -> None:
pass
def _is_consistent_lines(
self,
curr_line,
prev_line,
next_line,
consistent_direction, # 0 for prev, 1 for next, 2 for both
):
"""
This function checks if the line is consistent with its neighbors
Parameters
----------
curr_line : dict
current line
prev_line : dict
previous line
next_line : dict
next line
consistent_direction : int
0 for prev, 1 for next, 2 for both
Returns
-------
bool
True if the line is consistent with its neighbors, False otherwise.
"""
curr_line_font_size = curr_line["spans"][0]["size"]
curr_line_font_type = curr_line["spans"][0]["font"].lower()
if consistent_direction == 0:
if prev_line:
prev_line_font_size = prev_line["spans"][0]["size"]
prev_line_font_type = prev_line["spans"][0]["font"].lower()
return curr_line_font_size == prev_line_font_size and curr_line_font_type == prev_line_font_type
else:
return False
elif consistent_direction == 1:
if next_line:
next_line_font_size = next_line["spans"][0]["size"]
next_line_font_type = next_line["spans"][0]["font"].lower()
return curr_line_font_size == next_line_font_size and curr_line_font_type == next_line_font_type
else:
return False
elif consistent_direction == 2:
if prev_line and next_line:
prev_line_font_size = prev_line["spans"][0]["size"]
prev_line_font_type = prev_line["spans"][0]["font"].lower()
next_line_font_size = next_line["spans"][0]["size"]
next_line_font_type = next_line["spans"][0]["font"].lower()
return (curr_line_font_size == prev_line_font_size and curr_line_font_type == prev_line_font_type) and (
curr_line_font_size == next_line_font_size and curr_line_font_type == next_line_font_type
)
else:
return False
else:
return False
def _is_regular_line(self, curr_line_bbox, prev_line_bbox, next_line_bbox, avg_char_width, X0, X1, avg_line_height):
"""
This function checks if the line is a regular line
Parameters
----------
curr_line_bbox : list
bbox of the current line
prev_line_bbox : list
bbox of the previous line
next_line_bbox : list
bbox of the next line
avg_char_width : float
average of char widths
X0 : float
median of x0 values, which represents the left average boundary of the page
X1 : float
median of x1 values, which represents the right average boundary of the page
avg_line_height : float
average of line heights
Returns
-------
bool
True if the line is a regular line, False otherwise.
"""
horizontal_ratio = 0.5
vertical_ratio = 0.5
horizontal_thres = horizontal_ratio * avg_char_width
vertical_thres = vertical_ratio * avg_line_height
x0, y0, x1, y1 = curr_line_bbox
x0_near_X0 = abs(x0 - X0) < horizontal_thres
x1_near_X1 = abs(x1 - X1) < horizontal_thres
prev_line_is_end_of_para = prev_line_bbox and (abs(prev_line_bbox[2] - X1) > avg_char_width)
sufficient_spacing_above = False
if prev_line_bbox:
vertical_spacing_above = y1 - prev_line_bbox[3]
sufficient_spacing_above = vertical_spacing_above > vertical_thres
sufficient_spacing_below = False
if next_line_bbox:
vertical_spacing_below = next_line_bbox[1] - y0
sufficient_spacing_below = vertical_spacing_below > vertical_thres
return (
(sufficient_spacing_above or sufficient_spacing_below)
or (not x0_near_X0 and not x1_near_X1)
or prev_line_is_end_of_para
)
def _is_possible_start_of_para(self, curr_line, prev_line, next_line, X0, X1, avg_char_width, avg_font_size):
"""
This function checks if the line is a possible start of a paragraph
Parameters
----------
curr_line : dict
current line
prev_line : dict
previous line
next_line : dict
next line
X0 : float
median of x0 values, which represents the left average boundary of the page
X1 : float
median of x1 values, which represents the right average boundary of the page
avg_char_width : float
average of char widths
avg_line_height : float
average of line heights
Returns
-------
bool
True if the line is a possible start of a paragraph, False otherwise.
"""
start_confidence = 0.5 # Initial confidence of the line being a start of a paragraph
decision_path = [] # Record the decision path
curr_line_bbox = curr_line["bbox"]
prev_line_bbox = prev_line["bbox"] if prev_line else None
next_line_bbox = next_line["bbox"] if next_line else None
indent_ratio = 1
vertical_ratio = 1.5
vertical_thres = vertical_ratio * avg_font_size
left_horizontal_ratio = 0.5
left_horizontal_thres = left_horizontal_ratio * avg_char_width
right_horizontal_ratio = 2.5
right_horizontal_thres = right_horizontal_ratio * avg_char_width
x0, y0, x1, y1 = curr_line_bbox
indent_condition = x0 > X0 + indent_ratio * avg_char_width
if indent_condition:
start_confidence += 0.2
decision_path.append("indent_condition_met")
x0_near_X0 = abs(x0 - X0) < left_horizontal_thres
if x0_near_X0:
start_confidence += 0.1
decision_path.append("x0_near_X0")
x1_near_X1 = abs(x1 - X1) < right_horizontal_thres
if x1_near_X1:
start_confidence += 0.1
decision_path.append("x1_near_X1")
if prev_line is None:
prev_line_is_end_of_para = True
start_confidence += 0.2
decision_path.append("no_prev_line")
else:
prev_line_is_end_of_para, _, _ = self._is_possible_end_of_para(prev_line, next_line, X0, X1, avg_char_width)
if prev_line_is_end_of_para:
start_confidence += 0.1
decision_path.append("prev_line_is_end_of_para")
sufficient_spacing_above = False
if prev_line_bbox:
vertical_spacing_above = y1 - prev_line_bbox[3]
sufficient_spacing_above = vertical_spacing_above > vertical_thres
if sufficient_spacing_above:
start_confidence += 0.2
decision_path.append("sufficient_spacing_above")
sufficient_spacing_below = False
if next_line_bbox:
vertical_spacing_below = next_line_bbox[1] - y0
sufficient_spacing_below = vertical_spacing_below > vertical_thres
if sufficient_spacing_below:
start_confidence += 0.2
decision_path.append("sufficient_spacing_below")
is_regular_line = self._is_regular_line(
curr_line_bbox, prev_line_bbox, next_line_bbox, avg_char_width, X0, X1, avg_font_size
)
if is_regular_line:
start_confidence += 0.1
decision_path.append("is_regular_line")
is_start_of_para = (
(sufficient_spacing_above or sufficient_spacing_below)
or (indent_condition)
or (not indent_condition and x0_near_X0 and x1_near_X1 and not is_regular_line)
or prev_line_is_end_of_para
)
return (is_start_of_para, start_confidence, decision_path)
def _is_possible_end_of_para(self, curr_line, next_line, X0, X1, avg_char_width):
"""
This function checks if the line is a possible end of a paragraph
Parameters
----------
curr_line : dict
current line
next_line : dict
next line
X0 : float
median of x0 values, which represents the left average boundary of the page
X1 : float
median of x1 values, which represents the right average boundary of the page
avg_char_width : float
average of char widths
Returns
-------
bool
True if the line is a possible end of a paragraph, False otherwise.
"""
end_confidence = 0.5 # Initial confidence of the line being a end of a paragraph
decision_path = [] # Record the decision path
curr_line_bbox = curr_line["bbox"]
next_line_bbox = next_line["bbox"] if next_line else None
left_horizontal_ratio = 0.5
right_horizontal_ratio = 0.5
x0, _, x1, y1 = curr_line_bbox
next_x0, next_y0, _, _ = next_line_bbox if next_line_bbox else (0, 0, 0, 0)
x0_near_X0 = abs(x0 - X0) < left_horizontal_ratio * avg_char_width
if x0_near_X0:
end_confidence += 0.1
decision_path.append("x0_near_X0")
x1_smaller_than_X1 = x1 < X1 - right_horizontal_ratio * avg_char_width
if x1_smaller_than_X1:
end_confidence += 0.1
decision_path.append("x1_smaller_than_X1")
next_line_is_start_of_para = (
next_line_bbox
and (next_x0 > X0 + left_horizontal_ratio * avg_char_width)
and (not is_line_left_aligned_from_neighbors(curr_line_bbox, None, next_line_bbox, avg_char_width, direction=1))
)
if next_line_is_start_of_para:
end_confidence += 0.2
decision_path.append("next_line_is_start_of_para")
is_line_left_aligned_from_neighbors_bool = is_line_left_aligned_from_neighbors(
curr_line_bbox, None, next_line_bbox, avg_char_width
)
if is_line_left_aligned_from_neighbors_bool:
end_confidence += 0.1
decision_path.append("line_is_left_aligned_from_neighbors")
is_line_right_aligned_from_neighbors_bool = is_line_right_aligned_from_neighbors(
curr_line_bbox, None, next_line_bbox, avg_char_width
)
if not is_line_right_aligned_from_neighbors_bool:
end_confidence += 0.1
decision_path.append("line_is_not_right_aligned_from_neighbors")
is_end_of_para = end_with_punctuation(curr_line["text"]) and (
(x0_near_X0 and x1_smaller_than_X1)
or (is_line_left_aligned_from_neighbors_bool and not is_line_right_aligned_from_neighbors_bool)
)
return (is_end_of_para, end_confidence, decision_path)
def _cut_paras_per_block(
self,
block,
):
"""
Processes a raw block from PyMuPDF and returns the processed block.
Parameters
----------
raw_block : dict
A raw block from pymupdf.
Returns
-------
processed_block : dict
"""
def _construct_para(lines, is_block_title, para_title_level):
"""
Construct a paragraph from given lines.
"""
font_sizes = [span["size"] for line in lines for span in line["spans"]]
avg_font_size = sum(font_sizes) / len(font_sizes) if font_sizes else 0
font_colors = [span["color"] for line in lines for span in line["spans"]]
most_common_font_color = max(set(font_colors), key=font_colors.count) if font_colors else None
font_type_lengths = {}
for line in lines:
for span in line["spans"]:
font_type = span["font"]
bbox_width = span["bbox"][2] - span["bbox"][0]
if font_type in font_type_lengths:
font_type_lengths[font_type] += bbox_width
else:
font_type_lengths[font_type] = bbox_width
# get the font type with the longest bbox width
most_common_font_type = max(font_type_lengths, key=font_type_lengths.get) if font_type_lengths else None # type: ignore
para_bbox = calculate_para_bbox(lines)
para_text = " ".join(line["text"] for line in lines)
return {
"para_bbox": para_bbox,
"para_text": para_text,
"para_font_type": most_common_font_type,
"para_font_size": avg_font_size,
"para_font_color": most_common_font_color,
"is_para_title": is_block_title,
"para_title_level": para_title_level,
}
block_bbox = block["bbox"]
block_text = block["text"]
block_lines = block["lines"]
X0 = safe_get(block, "X0", 0)
X1 = safe_get(block, "X1", 0)
avg_char_width = safe_get(block, "avg_char_width", 0)
avg_char_height = safe_get(block, "avg_char_height", 0)
avg_font_size = safe_get(block, "avg_font_size", 0)
is_block_title = safe_get(block, "is_block_title", False)
para_title_level = safe_get(block, "block_title_level", 0)
# Segment into paragraphs
para_ranges = []
in_paragraph = False
start_idx_of_para = None
# Create the processed paragraphs
processed_paras = {}
para_bboxes = []
end_idx_of_para = 0
for line_index, line in enumerate(block_lines):
curr_line = line
prev_line = block_lines[line_index - 1] if line_index > 0 else None
next_line = block_lines[line_index + 1] if line_index < len(block_lines) - 1 else None
"""
Start processing paragraphs.
"""
# Check if the line is the start of a paragraph
is_start_of_para, start_confidence, decision_path = self._is_possible_start_of_para(
curr_line, prev_line, next_line, X0, X1, avg_char_width, avg_font_size
)
if not in_paragraph and is_start_of_para:
in_paragraph = True
start_idx_of_para = line_index
# print_green(">>> Start of a paragraph")
# print(" curr_line_text: ", curr_line["text"])
# print(" start_confidence: ", start_confidence)
# print(" decision_path: ", decision_path)
# Check if the line is the end of a paragraph
is_end_of_para, end_confidence, decision_path = self._is_possible_end_of_para(
curr_line, next_line, X0, X1, avg_char_width
)
if in_paragraph and (is_end_of_para or not next_line):
para_ranges.append((start_idx_of_para, line_index))
start_idx_of_para = None
in_paragraph = False
# print_red(">>> End of a paragraph")
# print(" curr_line_text: ", curr_line["text"])
# print(" end_confidence: ", end_confidence)
# print(" decision_path: ", decision_path)
# Add the last paragraph if it is not added
if in_paragraph and start_idx_of_para is not None:
para_ranges.append((start_idx_of_para, len(block_lines) - 1))
# Process the matched paragraphs
for para_index, (start_idx, end_idx) in enumerate(para_ranges):
matched_lines = block_lines[start_idx : end_idx + 1]
para_properties = _construct_para(matched_lines, is_block_title, para_title_level)
para_key = f"para_{len(processed_paras)}"
processed_paras[para_key] = para_properties
para_bboxes.append(para_properties["para_bbox"])
end_idx_of_para = end_idx + 1
# Deal with the remaining lines
if end_idx_of_para < len(block_lines):
unmatched_lines = block_lines[end_idx_of_para:]
unmatched_properties = _construct_para(unmatched_lines, is_block_title, para_title_level)
unmatched_key = f"para_{len(processed_paras)}"
processed_paras[unmatched_key] = unmatched_properties
para_bboxes.append(unmatched_properties["para_bbox"])
block["paras"] = processed_paras
return block
def batch_process_blocks(self, pdf_dict):
"""
Parses the blocks of all pages.
Parameters
----------
pdf_dict : dict
PDF dictionary.
filter_blocks : list
List of bounding boxes to filter.
Returns
-------
result_dict : dict
Result dictionary.
"""
num_paras = 0
for page_id, page in pdf_dict.items():
if page_id.startswith("page_"):
para_blocks = []
if "para_blocks" in page.keys():
input_blocks = page["para_blocks"]
for input_block in input_blocks:
new_block = self._cut_paras_per_block(input_block)
para_blocks.append(new_block)
num_paras += len(new_block["paras"])
page["para_blocks"] = para_blocks
pdf_dict["statistics"]["num_paras"] = num_paras
return pdf_dict
class BlockContinuationProcessor:
"""
This class is used to process the blocks to detect block continuations.
"""
def __init__(self) -> None:
pass
def __is_similar_font_type(self, font_type_1, font_type_2, prefix_length_ratio=0.3):
"""
This function checks if the two font types are similar.
Definition of similar font types: the two font types have a common prefix,
and the length of the common prefix is at least a certain ratio of the length of the shorter font type.
Parameters
----------
font_type1 : str
font type 1
font_type2 : str
font type 2
prefix_length_ratio : float
minimum ratio of the common prefix length to the length of the shorter font type
Returns
-------
bool
True if the two font types are similar, False otherwise.
"""
if isinstance(font_type_1, list):
font_type_1 = font_type_1[0] if font_type_1 else ""
if isinstance(font_type_2, list):
font_type_2 = font_type_2[0] if font_type_2 else ""
if font_type_1 == font_type_2:
return True
# Find the length of the common prefix
common_prefix_length = len(os.path.commonprefix([font_type_1, font_type_2]))
# Calculate the minimum prefix length based on the ratio
min_prefix_length = int(min(len(font_type_1), len(font_type_2)) * prefix_length_ratio)
return common_prefix_length >= min_prefix_length
def __is_same_block_font(self, block_1, block_2):
"""
This function compares the font of block1 and block2
Parameters
----------
block1 : dict
block1
block2 : dict
block2
Returns
-------
is_same : bool
True if block1 and block2 have the same font, else False
"""
block_1_font_type = safe_get(block_1, "block_font_type", "")
block_1_font_size = safe_get(block_1, "block_font_size", 0)
block_1_avg_char_width = safe_get(block_1, "avg_char_width", 0)
block_2_font_type = safe_get(block_2, "block_font_type", "")
block_2_font_size = safe_get(block_2, "block_font_size", 0)
block_2_avg_char_width = safe_get(block_2, "avg_char_width", 0)
if isinstance(block_1_font_size, list):
block_1_font_size = block_1_font_size[0] if block_1_font_size else 0
if isinstance(block_2_font_size, list):
block_2_font_size = block_2_font_size[0] if block_2_font_size else 0
block_1_text = safe_get(block_1, "text", "")
block_2_text = safe_get(block_2, "text", "")
if block_1_avg_char_width == 0 or block_2_avg_char_width == 0:
return False
if not block_1_text or not block_2_text:
return False
else:
text_len_ratio = len(block_2_text) / len(block_1_text)
if text_len_ratio < 0.2:
avg_char_width_condition = (
abs(block_1_avg_char_width - block_2_avg_char_width) / min(block_1_avg_char_width, block_2_avg_char_width)
< 0.5
)
else:
avg_char_width_condition = (
abs(block_1_avg_char_width - block_2_avg_char_width) / min(block_1_avg_char_width, block_2_avg_char_width)
< 0.2
)
block_font_size_condition = abs(block_1_font_size - block_2_font_size) < 1
return (
self.__is_similar_font_type(block_1_font_type, block_2_font_type)
and avg_char_width_condition
and block_font_size_condition
)
def _is_alphabet_char(self, char):
if (char >= "\u0041" and char <= "\u005a") or (char >= "\u0061" and char <= "\u007a"):
return True
else:
return False
def _is_chinese_char(self, char):
if char >= "\u4e00" and char <= "\u9fa5":
return True
else:
return False
def _is_other_letter_char(self, char):
try:
cat = unicodedata.category(char)
if cat == "Lu" or cat == "Ll":
return not self._is_alphabet_char(char) and not self._is_chinese_char(char)
except TypeError:
print("The input to the function must be a single character.")
return False
def _is_year(self, s: str):
try:
number = int(s)
return 1900 <= number <= 2099
except ValueError:
return False
def _match_brackets(self, text):
# pattern = r"^[\(\)\[\]（）【】{}｛｝<>＜＞〔〕〘〙\"\'“”‘’]"
pattern = r"^[\(\)\]（）】{}｛｝>＞〕〙\"\'“”‘’]"
return bool(re.match(pattern, text))
def _is_para_font_consistent(self, para_1, para_2):
"""
This function compares the font of para1 and para2
Parameters
----------
para1 : dict
para1
para2 : dict
para2
Returns
-------
is_same : bool
True if para1 and para2 have the same font, else False
"""
if para_1 is None or para_2 is None:
return False
para_1_font_type = safe_get(para_1, "para_font_type", "")
para_1_font_size = safe_get(para_1, "para_font_size", 0)
para_1_font_color = safe_get(para_1, "para_font_color", "")
para_2_font_type = safe_get(para_2, "para_font_type", "")
para_2_font_size = safe_get(para_2, "para_font_size", 0)
para_2_font_color = safe_get(para_2, "para_font_color", "")
if isinstance(para_1_font_type, list): # get the most common font type
para_1_font_type = max(set(para_1_font_type), key=para_1_font_type.count)
if isinstance(para_2_font_type, list):
para_2_font_type = max(set(para_2_font_type), key=para_2_font_type.count)
if isinstance(para_1_font_size, list): # compute average font type
para_1_font_size = sum(para_1_font_size) / len(para_1_font_size)
if isinstance(para_2_font_size, list): # compute average font type
para_2_font_size = sum(para_2_font_size) / len(para_2_font_size)
return (
self.__is_similar_font_type(para_1_font_type, para_2_font_type)
and abs(para_1_font_size - para_2_font_size) < 1.5
# and para_font_color1 == para_font_color2
)
def _is_para_puncs_consistent(self, para_1, para_2):
"""
This function determines whether para1 and para2 are originally from the same paragraph by checking the puncs of para1(former) and para2(latter)
Parameters
----------
para1 : dict
para1
para2 : dict
para2
Returns
-------
is_same : bool
True if para1 and para2 are from the same paragraph by using the puncs, else False
"""
para_1_text = safe_get(para_1, "para_text", "").strip()
para_2_text = safe_get(para_2, "para_text", "").strip()
para_1_bboxes = safe_get(para_1, "para_bbox", [])
para_1_font_sizes = safe_get(para_1, "para_font_size", 0)
para_2_bboxes = safe_get(para_2, "para_bbox", [])
para_2_font_sizes = safe_get(para_2, "para_font_size", 0)
# print_yellow(" Features of determine puncs_consistent:")
# print(f" para_1_text: {para_1_text}")
# print(f" para_2_text: {para_2_text}")
# print(f" para_1_bboxes: {para_1_bboxes}")
# print(f" para_2_bboxes: {para_2_bboxes}")
# print(f" para_1_font_sizes: {para_1_font_sizes}")
# print(f" para_2_font_sizes: {para_2_font_sizes}")
if is_nested_list(para_1_bboxes):
x0_1, y0_1, x1_1, y1_1 = para_1_bboxes[-1]
else:
x0_1, y0_1, x1_1, y1_1 = para_1_bboxes
if is_nested_list(para_2_bboxes):
x0_2, y0_2, x1_2, y1_2 = para_2_bboxes[0]
para_2_font_sizes = para_2_font_sizes[0] # type: ignore
else:
x0_2, y0_2, x1_2, y1_2 = para_2_bboxes
right_align_threshold = 0.5 * (para_1_font_sizes + para_2_font_sizes) * 0.8
are_two_paras_right_aligned = abs(x1_1 - x1_2) < right_align_threshold
left_indent_threshold = 0.5 * (para_1_font_sizes + para_2_font_sizes) * 0.8
is_para1_left_indent_than_papa2 = x0_1 - x0_2 > left_indent_threshold
is_para2_left_indent_than_papa1 = x0_2 - x0_1 > left_indent_threshold
# Check if either para_text1 or para_text2 is empty
if not para_1_text or not para_2_text:
return False
# Define the end puncs for a sentence to end and hyphen
end_puncs = [".", "?", "!", "。", "？", "！", "…"]
hyphen = ["-", "—"]
# Check if para_text1 ends with either hyphen or non-end punctuation or spaces
para_1_end_with_hyphen = para_1_text and para_1_text[-1] in hyphen
para_1_end_with_end_punc = para_1_text and para_1_text[-1] in end_puncs
para_1_end_with_space = para_1_text and para_1_text[-1] == " "
para_1_not_end_with_end_punc = para_1_text and para_1_text[-1] not in end_puncs
# print_yellow(f" para_1_end_with_hyphen: {para_1_end_with_hyphen}")
# print_yellow(f" para_1_end_with_end_punc: {para_1_end_with_end_punc}")
# print_yellow(f" para_1_not_end_with_end_punc: {para_1_not_end_with_end_punc}")
# print_yellow(f" para_1_end_with_space: {para_1_end_with_space}")
if para_1_end_with_hyphen: # If para_text1 ends with hyphen
# print_red(f"para_1 is end with hyphen.")
para_2_is_consistent = para_2_text and (
para_2_text[0] in hyphen
or (self._is_alphabet_char(para_2_text[0]) and para_2_text[0].islower())
or (self._is_chinese_char(para_2_text[0]))
or (self._is_other_letter_char(para_2_text[0]))
)
if para_2_is_consistent:
# print(f"para_2 is consistent.\n")
return True
else:
# print(f"para_2 is not consistent.\n")
pass
elif para_1_end_with_end_punc: # If para_text1 ends with ending punctuations
# print_red(f"para_1 is end with end_punc.")
para_2_is_consistent = (
para_2_text
and (
para_2_text[0]
== " "
# or (self._is_alphabet_char(para_2_text[0]) and para_2_text[0].isupper())
# or (self._is_chinese_char(para_2_text[0]))
# or (self._is_other_letter_char(para_2_text[0]))
)
and not is_para2_left_indent_than_papa1
)
if para_2_is_consistent:
# print(f"para_2 is consistent.\n")
return True
else:
# print(f"para_2 is not consistent.\n")
pass
elif para_1_not_end_with_end_punc: # If para_text1 is not end with ending punctuations
# print_red(f"para_1 is NOT end with end_punc.")
para_2_is_consistent = para_2_text and (
para_2_text[0] == " "
or (self._is_alphabet_char(para_2_text[0]) and para_2_text[0].islower())
or (self._is_alphabet_char(para_2_text[0]))
or (self._is_year(para_2_text[0:4]))
or (are_two_paras_right_aligned or is_para1_left_indent_than_papa2)
or (self._is_chinese_char(para_2_text[0]))
or (self._is_other_letter_char(para_2_text[0]))
or (self._match_brackets(para_2_text[0]))
)
if para_2_is_consistent:
# print(f"para_2 is consistent.\n")
return True
else:
# print(f"para_2 is not consistent.\n")
pass
elif para_1_end_with_space: # If para_text1 ends with space
# print_red(f"para_1 is end with space.")
para_2_is_consistent = para_2_text and (
para_2_text[0] == " "
or (self._is_alphabet_char(para_2_text[0]) and para_2_text[0].islower())
or (self._is_chinese_char(para_2_text[0]))
or (self._is_other_letter_char(para_2_text[0]))
)
if para_2_is_consistent:
# print(f"para_2 is consistent.\n")
return True
else:
pass
# print(f"para_2 is not consistent.\n")
return False
def _is_block_consistent(self, block_1, block_2):
"""
This function determines whether block1 and block2 are originally from the same block
Parameters
----------
block1 : dict
block1s
block2 : dict
block2
Returns
-------
is_same : bool
True if block1 and block2 are from the same block, else False
"""
return self.__is_same_block_font(block_1, block_2)
def _is_para_continued(self, para_1, para_2):
"""
This function determines whether para1 and para2 are originally from the same paragraph
Parameters
----------
para1 : dict
para1
para2 : dict
para2
Returns
-------
is_same : bool
True if para1 and para2 are from the same paragraph, else False
"""
is_para_font_consistent = self._is_para_font_consistent(para_1, para_2)
is_para_puncs_consistent = self._is_para_puncs_consistent(para_1, para_2)
return is_para_font_consistent and is_para_puncs_consistent
def _are_boundaries_of_block_consistent(self, block_1, block_2):
"""
This function checks if the boundaries of block1 and block2 are consistent
Parameters
----------
block1 : dict
block1
block2 : dict
block2
Returns
-------
is_consistent : bool
True if the boundaries of block1 and block2 are consistent, else False
"""
last_line_of_block_1 = block_1["lines"][-1]
first_line_of_block_2 = block_2["lines"][0]
spans_of_last_line_of_block_1 = last_line_of_block_1["spans"]
spans_of_first_line_of_block_2 = first_line_of_block_2["spans"]
font_type_of_last_line_of_block_1 = spans_of_last_line_of_block_1[0]["font"].lower()
font_size_of_last_line_of_block_1 = spans_of_last_line_of_block_1[0]["size"]
font_color_of_last_line_of_block_1 = spans_of_last_line_of_block_1[0]["color"]
font_flags_of_last_line_of_block_1 = spans_of_last_line_of_block_1[0]["flags"]
font_type_of_first_line_of_block_2 = spans_of_first_line_of_block_2[0]["font"].lower()
font_size_of_first_line_of_block_2 = spans_of_first_line_of_block_2[0]["size"]
font_color_of_first_line_of_block_2 = spans_of_first_line_of_block_2[0]["color"]
font_flags_of_first_line_of_block_2 = spans_of_first_line_of_block_2[0]["flags"]
return (
self.__is_similar_font_type(font_type_of_last_line_of_block_1, font_type_of_first_line_of_block_2)
and abs(font_size_of_last_line_of_block_1 - font_size_of_first_line_of_block_2) < 1
# and font_color_of_last_line_of_block1 == font_color_of_first_line_of_block2
and font_flags_of_last_line_of_block_1 == font_flags_of_first_line_of_block_2
)
def should_merge_next_para(self, curr_para, next_para):
"""
This function checks if the next_para should be merged into the curr_para.
Parameters
----------
curr_para : dict
The current paragraph.
next_para : dict
The next paragraph.
Returns
-------
bool
True if the next_para should be merged into the curr_para, False otherwise.
"""
if self._is_para_continued(curr_para, next_para):
return True
else:
return False
def batch_tag_paras(self, pdf_dict):
"""
This function tags the paragraphs in the pdf_dict.
Parameters
----------
pdf_dict : dict
PDF dictionary.
Returns
-------
pdf_dict : dict
PDF dictionary with tagged paragraphs.
"""
the_last_page_id = len(pdf_dict) - 1
for curr_page_idx, (curr_page_id, curr_page_content) in enumerate(pdf_dict.items()):
if curr_page_id.startswith("page_") and curr_page_content.get("para_blocks", []):
para_blocks_of_curr_page = curr_page_content["para_blocks"]
next_page_idx = curr_page_idx + 1
next_page_id = f"page_{next_page_idx}"
next_page_content = pdf_dict.get(next_page_id, {})
for i, current_block in enumerate(para_blocks_of_curr_page):
for para_id, curr_para in current_block["paras"].items():
curr_para["curr_para_location"] = [
curr_page_idx,
current_block["block_id"],
int(para_id.split("_")[-1]),
]
curr_para["next_para_location"] = None # 默认设置为None
curr_para["merge_next_para"] = False # 默认设置为False
next_block = para_blocks_of_curr_page[i + 1] if i < len(para_blocks_of_curr_page) - 1 else None
if next_block:
curr_block_last_para_key = list(current_block["paras"].keys())[-1]
curr_blk_last_para = current_block["paras"][curr_block_last_para_key]
next_block_first_para_key = list(next_block["paras"].keys())[0]
next_blk_first_para = next_block["paras"][next_block_first_para_key]
if self.should_merge_next_para(curr_blk_last_para, next_blk_first_para):
curr_blk_last_para["next_para_location"] = [
curr_page_idx,
next_block["block_id"],
int(next_block_first_para_key.split("_")[-1]),
]
curr_blk_last_para["merge_next_para"] = True
else:
# Handle the case where the next block is in a different page
curr_block_last_para_key = list(current_block["paras"].keys())[-1]
curr_blk_last_para = current_block["paras"][curr_block_last_para_key]
while not next_page_content.get("para_blocks", []) and next_page_idx <= the_last_page_id:
next_page_idx += 1
next_page_id = f"page_{next_page_idx}"
next_page_content = pdf_dict.get(next_page_id, {})
if next_page_content.get("para_blocks", []):
next_blk_first_para_key = list(next_page_content["para_blocks"][0]["paras"].keys())[0]
next_blk_first_para = next_page_content["para_blocks"][0]["paras"][next_blk_first_para_key]
if self.should_merge_next_para(curr_blk_last_para, next_blk_first_para):
curr_blk_last_para["next_para_location"] = [
next_page_idx,
next_page_content["para_blocks"][0]["block_id"],
int(next_blk_first_para_key.split("_")[-1]),
]
curr_blk_last_para["merge_next_para"] = True
return pdf_dict
def find_block_by_id(self, para_blocks, block_id):
"""
This function finds a block by its id.
Parameters
----------
para_blocks : list
List of blocks.
block_id : int
Id of the block to find.
Returns
-------
block : dict
The block with the given id.
"""
for blk_idx, block in enumerate(para_blocks):
if block.get("block_id") == block_id:
return block
return None
def batch_merge_paras(self, pdf_dict):
"""
This function merges the paragraphs in the pdf_dict.
Parameters
----------
pdf_dict : dict
PDF dictionary.
Returns
-------
pdf_dict : dict
PDF dictionary with merged paragraphs.
"""
for page_id, page_content in pdf_dict.items():
if page_id.startswith("page_") and page_content.get("para_blocks", []):
para_blocks_of_page = page_content["para_blocks"]
for i in range(len(para_blocks_of_page)):
current_block = para_blocks_of_page[i]
paras = current_block["paras"]
for para_id, curr_para in list(paras.items()):
# print(f"current para_id: {para_id}")
# 跳过标题段落
if curr_para.get("is_para_title"):
continue
while curr_para.get("merge_next_para"):
curr_para_location = curr_para.get("curr_para_location")
next_para_location = curr_para.get("next_para_location")
# print(f"curr_para_location: {curr_para_location}, next_para_location: {next_para_location}")
if not next_para_location:
break
if curr_para_location == next_para_location:
# print_red("The next para is in the same block as the current para.")
curr_para["merge_next_para"] = False
break
next_page_idx, next_block_id, next_para_id = next_para_location
next_page_id = f"page_{next_page_idx}"
next_page_content = pdf_dict.get(next_page_id)
if not next_page_content:
break
next_block = self.find_block_by_id(next_page_content.get("para_blocks", []), next_block_id)
if not next_block:
break
next_para = next_block["paras"].get(f"para_{next_para_id}")
if not next_para or next_para.get("is_para_title"):
break
# 合并段落文本
curr_para_text = curr_para.get("para_text", "")
next_para_text = next_para.get("para_text", "")
curr_para["para_text"] = curr_para_text + " " + next_para_text
# 更新 next_para_location
curr_para["next_para_location"] = next_para.get("next_para_location")
# 将下一个段落文本置为空，表示已被合并
next_para["para_text"] = ""
# 更新 merge_next_para 标记
curr_para["merge_next_para"] = next_para.get("merge_next_para", False)
return pdf_dict
class DrawAnnos:
"""
This class draws annotations on the pdf file
----------------------------------------
Color Code
----------------------------------------
Red: (1, 0, 0)
Green: (0, 1, 0)
Blue: (0, 0, 1)
Yellow: (1, 1, 0) - mix of red and green
Cyan: (0, 1, 1) - mix of green and blue
Magenta: (1, 0, 1) - mix of red and blue
White: (1, 1, 1) - red, green and blue full intensity
Black: (0, 0, 0) - no color component whatsoever
Gray: (0.5, 0.5, 0.5) - equal and medium intensity of red, green and blue color components
Orange: (1, 0.65, 0) - maximum intensity of red, medium intensity of green, no blue component
"""
def __init__(self) -> None:
pass
def __is_nested_list(self, lst):
"""
This function returns True if the given list is a nested list of any degree.
"""
if isinstance(lst, list):
return any(self.__is_nested_list(i) for i in lst) or any(isinstance(i, list) for i in lst)
return False
def __valid_rect(self, bbox):
# Ensure that the rectangle is not empty or invalid
if isinstance(bbox[0], list):
return False # It's a nested list, hence it can't be valid rect
else:
return bbox[0] < bbox[2] and bbox[1] < bbox[3]
def __draw_nested_boxes(self, page, nested_bbox, color=(0, 1, 1)):
"""
This function draws the nested boxes
Parameters
----------
page : fitz.Page
page
nested_bbox : list
nested bbox
color : tuple
color, by default (0, 1, 1) # draw with cyan color for combined paragraph
"""
if self.__is_nested_list(nested_bbox): # If it's a nested list
for bbox in nested_bbox:
self.__draw_nested_boxes(page, bbox, color) # Recursively call the function
elif self.__valid_rect(nested_bbox): # If valid rectangle
para_rect = fitz.Rect(nested_bbox)
para_anno = page.add_rect_annot(para_rect)
para_anno.set_colors(stroke=color) # draw with cyan color for combined paragraph
para_anno.set_border(width=1)
para_anno.update()
def draw_annos(self, input_pdf_path, pdf_dic, output_pdf_path):
"""
This function draws annotations on the pdf file.
Parameters
----------
input_pdf_path : str
path to the input pdf file
pdf_dic : dict
pdf dictionary
output_pdf_path : str
path to the output pdf file
pdf_dic : dict
pdf dictionary
"""
pdf_doc = open_pdf(input_pdf_path)
if pdf_dic is None:
pdf_dic = {}
if output_pdf_path is None:
output_pdf_path = input_pdf_path.replace(".pdf", "_anno.pdf")
for page_id, page in enumerate(pdf_doc): # type: ignore
page_key = f"page_{page_id}"
for ele_key, ele_data in pdf_dic[page_key].items():
if ele_key == "para_blocks":
para_blocks = ele_data
for para_block in para_blocks:
if "paras" in para_block.keys():
paras = para_block["paras"]
for para_key, para_content in paras.items():
para_bbox = para_content["para_bbox"]
# print(f"para_bbox: {para_bbox}")
# print(f"is a nested list: {self.__is_nested_list(para_bbox)}")
if self.__is_nested_list(para_bbox) and len(para_bbox) > 1:
color = (0, 1, 1)
self.__draw_nested_boxes(
page, para_bbox, color
) # draw with cyan color for combined paragraph
else:
if self.__valid_rect(para_bbox):
para_rect = fitz.Rect(para_bbox)
para_anno = page.add_rect_annot(para_rect)
para_anno.set_colors(stroke=(0, 1, 0)) # draw with green color for normal paragraph
para_anno.set_border(width=0.5)
para_anno.update()
is_para_title = para_content["is_para_title"]
if is_para_title:
if self.__is_nested_list(para_content["para_bbox"]) and len(para_content["para_bbox"]) > 1:
color = (0, 0, 1)
self.__draw_nested_boxes(
page, para_content["para_bbox"], color
) # draw with cyan color for combined title
else:
if self.__valid_rect(para_content["para_bbox"]):
para_rect = fitz.Rect(para_content["para_bbox"])
if self.__valid_rect(para_content["para_bbox"]):
para_anno = page.add_rect_annot(para_rect)
para_anno.set_colors(stroke=(0, 0, 1)) # draw with blue color for normal title
para_anno.set_border(width=0.5)
para_anno.update()
pdf_doc.save(output_pdf_path)
pdf_doc.close()
class ParaProcessPipeline:
def __init__(self) -> None:
pass
def para_process_pipeline(self, pdf_info_dict, para_debug_mode=None, input_pdf_path=None, output_pdf_path=None):
"""
This function processes the paragraphs, including:
1. Read raw input json file into pdf_dic
2. Detect and replace equations
3. Combine spans into a natural line
4. Check if the paragraphs are inside bboxes passed from "layout_bboxes" key
5. Compute statistics for each block
6. Detect titles in the document
7. Detect paragraphs inside each block
8. Divide the level of the titles
9. Detect and combine paragraphs from different blocks into one paragraph
10. Check whether the final results after checking headings, dividing paragraphs within blocks, and merging paragraphs between blocks are plausible and reasonable.
11. Draw annotations on the pdf file
Parameters
----------
pdf_dic_json_fpath : str
path to the pdf dictionary json file.
Notice: data noises, including overlap blocks, header, footer, watermark, vertical margin note have been removed already.
input_pdf_doc : str
path to the input pdf file
output_pdf_path : str
path to the output pdf file
Returns
-------
pdf_dict : dict
result dictionary
"""
error_info = None
output_json_file = ""
output_dir = ""
if input_pdf_path is not None:
input_pdf_path = os.path.abspath(input_pdf_path)
# print_green_on_red(f">>>>>>>>>>>>>>>>>>> Process the paragraphs of {input_pdf_path}")
if output_pdf_path is not None:
output_dir = os.path.dirname(output_pdf_path)
output_json_file = f"{output_dir}/pdf_dic.json"
def __save_pdf_dic(pdf_dic, output_pdf_path, stage="0", para_debug_mode=para_debug_mode):
"""
Save the pdf_dic to a json file
"""
output_pdf_file_name = os.path.basename(output_pdf_path)
# output_dir = os.path.dirname(output_pdf_path)
output_dir = "\\tmp\\pdf_parse"
output_pdf_file_name = output_pdf_file_name.replace(".pdf", f"_stage_{stage}.json")
pdf_dic_json_fpath = os.path.join(output_dir, output_pdf_file_name)
if not os.path.exists(output_dir):
os.makedirs(output_dir)
if para_debug_mode == "full":
with open(pdf_dic_json_fpath, "w", encoding="utf-8") as f:
json.dump(pdf_dic, f, indent=2, ensure_ascii=False)
# Validate the output already exists
if not os.path.exists(pdf_dic_json_fpath):
print_red(f"Failed to save the pdf_dic to {pdf_dic_json_fpath}")
return None
else:
print_green(f"Succeed to save the pdf_dic to {pdf_dic_json_fpath}")
return pdf_dic_json_fpath
"""
Preprocess the lines of block
"""
# Combine spans into a natural line
rawBlockProcessor = RawBlockProcessor()
pdf_dic = rawBlockProcessor.batch_process_blocks(pdf_info_dict)
# print(f"pdf_dic['page_0']['para_blocks'][0]: {pdf_dic['page_0']['para_blocks'][0]}", end="\n\n")
# Check if the paragraphs are inside bboxes passed from "layout_bboxes" key
layoutFilter = LayoutFilterProcessor()
pdf_dic = layoutFilter.batch_process_blocks(pdf_dic)
# Compute statistics for each block
blockStatisticsCalculator = BlockStatisticsCalculator()
pdf_dic = blockStatisticsCalculator.batch_process_blocks(pdf_dic)
# print(f"pdf_dic['page_0']['para_blocks'][0]: {pdf_dic['page_0']['para_blocks'][0]}", end="\n\n")
# Compute statistics for all blocks(namely this pdf document)
docStatisticsCalculator = DocStatisticsCalculator()
pdf_dic = docStatisticsCalculator.calc_stats_of_doc(pdf_dic)
# print(f"pdf_dic['statistics']: {pdf_dic['statistics']}", end="\n\n")
# Dump the first three stages of pdf_dic to a json file
if para_debug_mode == "full":
pdf_dic_json_fpath = __save_pdf_dic(pdf_dic, output_pdf_path, stage="0", para_debug_mode=para_debug_mode)
"""
Detect titles in the document
"""
doc_statistics = pdf_dic["statistics"]
titleProcessor = TitleProcessor(doc_statistics)
pdf_dic = titleProcessor.batch_detect_titles(pdf_dic)
if para_debug_mode == "full":
pdf_dic_json_fpath = __save_pdf_dic(pdf_dic, output_pdf_path, stage="1", para_debug_mode=para_debug_mode)
"""
Detect and divide the level of the titles
"""
titleProcessor = TitleProcessor()
pdf_dic = titleProcessor.batch_recog_title_level(pdf_dic)
if para_debug_mode == "full":
pdf_dic_json_fpath = __save_pdf_dic(pdf_dic, output_pdf_path, stage="2", para_debug_mode=para_debug_mode)
"""
Detect and split paragraphs inside each block
"""
blockInnerParasProcessor = BlockTerminationProcessor()
pdf_dic = blockInnerParasProcessor.batch_process_blocks(pdf_dic)
if para_debug_mode == "full":
pdf_dic_json_fpath = __save_pdf_dic(pdf_dic, output_pdf_path, stage="3", para_debug_mode=para_debug_mode)
# pdf_dic_json_fpath = __save_pdf_dic(pdf_dic, output_pdf_path, stage="3", para_debug_mode="full")
# print_green(f"pdf_dic_json_fpath: {pdf_dic_json_fpath}")
"""
Detect and combine paragraphs from different blocks into one paragraph
"""
blockContinuationProcessor = BlockContinuationProcessor()
pdf_dic = blockContinuationProcessor.batch_tag_paras(pdf_dic)
pdf_dic = blockContinuationProcessor.batch_merge_paras(pdf_dic)
if para_debug_mode == "full":
pdf_dic_json_fpath = __save_pdf_dic(pdf_dic, output_pdf_path, stage="4", para_debug_mode=para_debug_mode)
# pdf_dic_json_fpath = __save_pdf_dic(pdf_dic, output_pdf_path, stage="4", para_debug_mode="full")
# print_green(f"pdf_dic_json_fpath: {pdf_dic_json_fpath}")
"""
Discard pdf files by checking exceptions and return the error info to the caller
"""
discardByException = DiscardByException()
is_discard_by_single_line_block = discardByException.discard_by_single_line_block(
pdf_dic, exception=DenseSingleLineBlockException()
)
is_discard_by_title_detection = discardByException.discard_by_title_detection(
pdf_dic, exception=TitleDetectionException()
)
is_discard_by_title_level = discardByException.discard_by_title_level(pdf_dic, exception=TitleLevelException())
is_discard_by_split_para = discardByException.discard_by_split_para(pdf_dic, exception=ParaSplitException())
is_discard_by_merge_para = discardByException.discard_by_merge_para(pdf_dic, exception=ParaMergeException())
if is_discard_by_single_line_block is not None:
error_info = is_discard_by_single_line_block
elif is_discard_by_title_detection is not None:
error_info = is_discard_by_title_detection
elif is_discard_by_title_level is not None:
error_info = is_discard_by_title_level
elif is_discard_by_split_para is not None:
error_info = is_discard_by_split_para
elif is_discard_by_merge_para is not None:
error_info = is_discard_by_merge_para
if error_info is not None:
return pdf_dic, error_info
"""
Dump the final pdf_dic to a json file
"""
if para_debug_mode is not None:
with open(output_json_file, "w", encoding="utf-8") as f:
json.dump(pdf_info_dict, f, ensure_ascii=False, indent=4)
"""
Draw the annotations
"""
if para_debug_mode is not None:
drawAnnos = DrawAnnos()
drawAnnos.draw_annos(input_pdf_path, pdf_dic, output_pdf_path)
"""
Remove the intermediate files which are generated in the process of paragraph processing if debug_mode is simple
"""
if para_debug_mode is not None:
for fpath in os.listdir(output_dir):
if fpath.endswith(".json") and "stage" in fpath:
os.remove(os.path.join(output_dir, fpath))
return pdf_dic, error_info
"""
Run this script to test the function with Command:
python detect_para.py [pdf_path] [output_pdf_path]
Params:
- pdf_path: the path of the pdf file
- output_pdf_path: the path of the output pdf file
"""
if __name__ == "__main__":
DEFAULT_PDF_PATH = (
"app/pdf_toolbox/tests/assets/paper/paper.pdf" if os.name != "nt" else "app\\pdf_toolbox\\tests\\assets\\paper\\paper.pdf"
)
input_pdf_path = sys.argv[1] if len(sys.argv) > 1 else DEFAULT_PDF_PATH
output_pdf_path = sys.argv[2] if len(sys.argv) > 2 else input_pdf_path.split(".")[0] + "_recogPara.pdf"
output_json_path = sys.argv[3] if len(sys.argv) > 3 else input_pdf_path.split(".")[0] + "_recogPara.json"
import stat
# Remove existing output file if it exists
if os.path.exists(output_pdf_path):
os.chmod(output_pdf_path, stat.S_IWRITE)
os.remove(output_pdf_path)
input_pdf_doc = open_pdf(input_pdf_path)
# postprocess the paragraphs
paraProcessPipeline = ParaProcessPipeline()
# parse paragraph and save to json file
pdf_dic = {}
blockInnerParasProcessor = BlockTerminationProcessor()
"""
Construct the pdf dictionary.
"""
for page_id, page in enumerate(input_pdf_doc): # type: ignore
# print(f"Processing page {page_id}")
# print(f"page: {page}")
raw_blocks = page.get_text("dict")["blocks"]
# Save text blocks to "preproc_blocks"
preproc_blocks = []
for block in raw_blocks:
if block["type"] == 0:
preproc_blocks.append(block)
layout_bboxes = []
# Construct the pdf dictionary as schema above
page_dict = {
"para_blocks": None,
"preproc_blocks": preproc_blocks,
"images": None,
"tables": None,
"interline_equations": None,
"inline_equations": None,
"layout_bboxes": None,
"pymu_raw_blocks": None,
"global_statistic": None,
"droped_text_block": None,
"droped_image_block": None,
"droped_table_block": None,
"image_backup": None,
"table_backup": None,
}
pdf_dic[f"page_{page_id}"] = page_dict
# print(f"pdf_dic: {pdf_dic}")
with open(output_json_path, "w", encoding="utf-8") as f:
json.dump(pdf_dic, f, ensure_ascii=False, indent=4)
pdf_dic = paraProcessPipeline.para_process_pipeline(output_json_path, input_pdf_doc, output_pdf_path)
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment