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
Hide whitespace changes
Inline Side-by-side
Showing with 13 additions and 3821 deletions
README.md
View file @ 8afff9ae
...@@ -277,88 +277,14 @@ If your device supports CUDA and meets the GPU requirements of the mainline envi ...@@ -277,88 +277,14 @@ If your device supports CUDA and meets the GPU requirements of the mainline envi
### Command Line ### Command Line
```bash [Using MinerU via Command Line](https://mineru.readthedocs.io/en/latest/user_guide/quick_start/command_line.html)
magic-pdf --help
Usage: magic-pdf [OPTIONS]
Options:
-v, --version display the version and exit
-p, --path PATH local pdf filepath or directory [required]
-o, --output-dir PATH output local directory [required]
-m, --method [ocr|txt|auto] the method for parsing pdf. ocr: using ocr
technique to extract information from pdf. txt:
suitable for the text-based pdf only and
outperform ocr. auto: automatically choose the
best method for parsing pdf from ocr and txt.
without method specified, auto will be used by
default.
-l, --lang TEXT Input the languages in the pdf (if known) to
improve OCR accuracy. Optional. You should
input "Abbreviation" with language form url: ht
tps://paddlepaddle.github.io/PaddleOCR/latest/en
/ppocr/blog/multi_languages.html#5-support-languages-
and-abbreviations
-d, --debug BOOLEAN Enables detailed debugging information during
the execution of the CLI commands.
-s, --start INTEGER The starting page for PDF parsing, beginning
from 0.
-e, --end INTEGER The ending page for PDF parsing, beginning from
0.
--help Show this message and exit.
## show version
magic-pdf -v
## command line example
magic-pdf -p {some_pdf} -o {some_output_dir} -m auto
```
`{some_pdf}` can be a single PDF file or a directory containing multiple PDFs.
The results will be saved in the `{some_output_dir}` directory. The output file list is as follows:
```text
├── some_pdf.md # markdown file
├── images # directory for storing images
├── some_pdf_layout.pdf # layout diagram (Include layout reading order)
├── some_pdf_middle.json # MinerU intermediate processing result
├── some_pdf_model.json # model inference result
├── some_pdf_origin.pdf # original PDF file
├── some_pdf_spans.pdf # smallest granularity bbox position information diagram
└── some_pdf_content_list.json # Rich text JSON arranged in reading order
```
> [!TIP] > [!TIP]
> For more information about the output files, please refer to the [Output File Description](docs/output_file_en_us.md). > For more information about the output files, please refer to the [Output File Description](docs/output_file_en_us.md).
### API ### API
Processing files from local disk [Using MinerU via Python API](https://mineru.readthedocs.io/en/latest/user_guide/quick_start/to_markdown.html)
```python
image_writer = DiskReaderWriter(local_image_dir)
image_dir = str(os.path.basename(local_image_dir))
jso_useful_key = {"_pdf_type": "", "model_list": []}
pipe = UNIPipe(pdf_bytes, jso_useful_key, image_writer)
pipe.pipe_classify()
pipe.pipe_analyze()
pipe.pipe_parse()
md_content = pipe.pipe_mk_markdown(image_dir, drop_mode="none")
```
Processing files from object storage
```python
s3pdf_cli = S3ReaderWriter(pdf_ak, pdf_sk, pdf_endpoint)
image_dir = "s3://img_bucket/"
s3image_cli = S3ReaderWriter(img_ak, img_sk, img_endpoint, parent_path=image_dir)
pdf_bytes = s3pdf_cli.read(s3_pdf_path, mode=s3pdf_cli.MODE_BIN)
jso_useful_key = {"_pdf_type": "", "model_list": []}
pipe = UNIPipe(pdf_bytes, jso_useful_key, s3image_cli)
pipe.pipe_classify()
pipe.pipe_analyze()
pipe.pipe_parse()
md_content = pipe.pipe_mk_markdown(image_dir, drop_mode="none")
```
For detailed implementation, refer to: For detailed implementation, refer to:
......
README_zh-CN.md
View file @ 8afff9ae
...@@ -284,89 +284,14 @@ pip install -U magic-pdf[full] --extra-index-url https://wheels.myhloli.com -i h ...@@ -284,89 +284,14 @@ pip install -U magic-pdf[full] --extra-index-url https://wheels.myhloli.com -i h
### 命令行 ### 命令行
```bash [通过命令行使用MinerU](https://mineru.readthedocs.io/zh-cn/latest/user_guide/quick_start/command_line.html)
magic-pdf --help
Usage: magic-pdf [OPTIONS]
Options:
-v, --version display the version and exit
-p, --path PATH local pdf filepath or directory [required]
-o, --output-dir PATH output local directory [required]
-m, --method [ocr|txt|auto] the method for parsing pdf. ocr: using ocr
technique to extract information from pdf. txt:
suitable for the text-based pdf only and
outperform ocr. auto: automatically choose the
best method for parsing pdf from ocr and txt.
without method specified, auto will be used by
default.
-l, --lang TEXT Input the languages in the pdf (if known) to
improve OCR accuracy. Optional. You should
input "Abbreviation" with language form url: ht
tps://paddlepaddle.github.io/PaddleOCR/latest/en
/ppocr/blog/multi_languages.html#5-support-languages-
and-abbreviations
-d, --debug BOOLEAN Enables detailed debugging information during
the execution of the CLI commands.
-s, --start INTEGER The starting page for PDF parsing, beginning
from 0.
-e, --end INTEGER The ending page for PDF parsing, beginning from
0.
--help Show this message and exit.
## show version
magic-pdf -v
## command line example
magic-pdf -p {some_pdf} -o {some_output_dir} -m auto
```
其中 `{some_pdf}` 可以是单个pdf文件，也可以是一个包含多个pdf文件的目录。
运行完命令后输出的结果会保存在`{some_output_dir}`目录下, 输出的文件列表如下
```text
├── some_pdf.md # markdown 文件
├── images # 存放图片目录
├── some_pdf_layout.pdf # layout 绘图 （包含layout阅读顺序）
├── some_pdf_middle.json # minerU 中间处理结果
├── some_pdf_model.json # 模型推理结果
├── some_pdf_origin.pdf # 原 pdf 文件
├── some_pdf_spans.pdf # 最小粒度的bbox位置信息绘图
└── some_pdf_content_list.json # 按阅读顺序排列的富文本json
```
> [!TIP] > [!TIP]
> 更多有关输出文件的信息，请参考[输出文件说明](docs/output_file_zh_cn.md) > 更多有关输出文件的信息，请参考[输出文件说明](docs/output_file_zh_cn.md)
### API ### API
处理本地磁盘上的文件 [通过Python代码调用MinerU](https://mineru.readthedocs.io/zh-cn/latest/user_guide/quick_start/to_markdown.html)
```python
image_writer = DiskReaderWriter(local_image_dir)
image_dir = str(os.path.basename(local_image_dir))
jso_useful_key = {"_pdf_type": "", "model_list": []}
pipe = UNIPipe(pdf_bytes, jso_useful_key, image_writer)
pipe.pipe_classify()
pipe.pipe_analyze()
pipe.pipe_parse()
md_content = pipe.pipe_mk_markdown(image_dir, drop_mode="none")
```
处理对象存储上的文件
```python
s3pdf_cli = S3ReaderWriter(pdf_ak, pdf_sk, pdf_endpoint)
image_dir = "s3://img_bucket/"
s3image_cli = S3ReaderWriter(img_ak, img_sk, img_endpoint, parent_path=image_dir)
pdf_bytes = s3pdf_cli.read(s3_pdf_path, mode=s3pdf_cli.MODE_BIN)
jso_useful_key = {"_pdf_type": "", "model_list": []}
pipe = UNIPipe(pdf_bytes, jso_useful_key, s3image_cli)
pipe.pipe_classify()
pipe.pipe_analyze()
pipe.pipe_parse()
md_content = pipe.pipe_mk_markdown(image_dir, drop_mode="none")
```
详细实现可参考 详细实现可参考
......
magic_pdf/dict2md/mkcontent.py deleted 100644 → 0
View file @ 4df1eb74
import math
from loguru import logger
from magic_pdf.config.ocr_content_type import ContentType
from magic_pdf.libs.boxbase import (find_bottom_nearest_text_bbox,
find_top_nearest_text_bbox)
from magic_pdf.libs.commons import join_path
TYPE_INLINE_EQUATION = ContentType.InlineEquation
TYPE_INTERLINE_EQUATION = ContentType.InterlineEquation
UNI_FORMAT_TEXT_TYPE = ['text', 'h1', 'h2', 'h3', 'h4', 'h5', 'h6']
@DeprecationWarning
def mk_nlp_markdown_1(para_dict: dict):
"""对排序后的bboxes拼接内容."""
content_lst = []
for _, page_info in para_dict.items():
para_blocks = page_info.get('para_blocks')
if not para_blocks:
continue
for block in para_blocks:
item = block['paras']
for _, p in item.items():
para_text = p['para_text']
is_title = p['is_para_title']
title_level = p['para_title_level']
md_title_prefix = '#' * title_level
if is_title:
content_lst.append(f'{md_title_prefix} {para_text}')
else:
content_lst.append(para_text)
content_text = '\n\n'.join(content_lst)
return content_text
# 找到目标字符串在段落中的索引
def __find_index(paragraph, target):
index = paragraph.find(target)
if index != -1:
return index
else:
return None
def __insert_string(paragraph, target, position):
new_paragraph = paragraph[:position] + target + paragraph[position:]
return new_paragraph
def __insert_after(content, image_content, target):
"""在content中找到target，将image_content插入到target后面."""
index = content.find(target)
if index != -1:
content = (
content[: index + len(target)]
+ '\n\n'
+ image_content
+ '\n\n'
+ content[index + len(target) :]
)
else:
logger.error(
f"Can't find the location of image {image_content} in the markdown file, search target is {target}"
)
return content
def __insert_before(content, image_content, target):
"""在content中找到target，将image_content插入到target前面."""
index = content.find(target)
if index != -1:
content = content[:index] + '\n\n' + image_content + '\n\n' + content[index:]
else:
logger.error(
f"Can't find the location of image {image_content} in the markdown file, search target is {target}"
)
return content
@DeprecationWarning
def mk_mm_markdown_1(para_dict: dict):
"""拼装多模态markdown."""
content_lst = []
for _, page_info in para_dict.items():
page_lst = [] # 一个page内的段落列表
para_blocks = page_info.get('para_blocks')
pymu_raw_blocks = page_info.get('preproc_blocks')
all_page_images = []
all_page_images.extend(page_info.get('images', []))
all_page_images.extend(page_info.get('image_backup', []))
all_page_images.extend(page_info.get('tables', []))
all_page_images.extend(page_info.get('table_backup', []))
if not para_blocks or not pymu_raw_blocks: # 只有图片的拼接的场景
for img in all_page_images:
page_lst.append(f"![]({img['image_path']})") # TODO 图片顺序
page_md = '\n\n'.join(page_lst)
else:
for block in para_blocks:
item = block['paras']
for _, p in item.items():
para_text = p['para_text']
is_title = p['is_para_title']
title_level = p['para_title_level']
md_title_prefix = '#' * title_level
if is_title:
page_lst.append(f'{md_title_prefix} {para_text}')
else:
page_lst.append(para_text)
"""拼装成一个页面的文本"""
page_md = '\n\n'.join(page_lst)
"""插入图片"""
for img in all_page_images:
imgbox = img['bbox']
img_content = f"![]({img['image_path']})"
# 先看在哪个block内
for block in pymu_raw_blocks:
bbox = block['bbox']
if (
bbox[0] - 1 <= imgbox[0] < bbox[2] + 1
and bbox[1] - 1 <= imgbox[1] < bbox[3] + 1
): # 确定在block内
for l in block['lines']: # noqa: E741
line_box = l['bbox']
if (
line_box[0] - 1 <= imgbox[0] < line_box[2] + 1
and line_box[1] - 1 <= imgbox[1] < line_box[3] + 1
): # 在line内的，插入line前面
line_txt = ''.join([s['text'] for s in l['spans']])
page_md = __insert_before(
page_md, img_content, line_txt
)
break
break
else: # 在行与行之间
# 找到图片x0,y0与line的x0,y0最近的line
min_distance = 100000
min_line = None
for l in block['lines']: # noqa: E741
line_box = l['bbox']
distance = math.sqrt(
(line_box[0] - imgbox[0]) ** 2
+ (line_box[1] - imgbox[1]) ** 2
)
if distance < min_distance:
min_distance = distance
min_line = l
if min_line:
line_txt = ''.join(
[s['text'] for s in min_line['spans']]
)
img_h = imgbox[3] - imgbox[1]
if min_distance < img_h: # 文字在图片前面
page_md = __insert_after(
page_md, img_content, line_txt
)
else:
page_md = __insert_before(
page_md, img_content, line_txt
)
else:
logger.error(
f"Can't find the location of image {img['image_path']} in the markdown file #1"
)
else: # 应当在两个block之间
# 找到上方最近的block，如果上方没有就找大下方最近的block
top_txt_block = find_top_nearest_text_bbox(pymu_raw_blocks, imgbox)
if top_txt_block:
line_txt = ''.join(
[s['text'] for s in top_txt_block['lines'][-1]['spans']]
)
page_md = __insert_after(page_md, img_content, line_txt)
else:
bottom_txt_block = find_bottom_nearest_text_bbox(
pymu_raw_blocks, imgbox
)
if bottom_txt_block:
line_txt = ''.join(
[
s['text']
for s in bottom_txt_block['lines'][0]['spans']
]
)
page_md = __insert_before(page_md, img_content, line_txt)
else:
logger.error(
f"Can't find the location of image {img['image_path']} in the markdown file #2"
)
content_lst.append(page_md)
"""拼装成全部页面的文本"""
content_text = '\n\n'.join(content_lst)
return content_text
def __insert_after_para(text, type, element, content_list):
"""在content_list中找到text，将image_path作为一个新的node插入到text后面."""
for i, c in enumerate(content_list):
content_type = c.get('type')
if content_type in UNI_FORMAT_TEXT_TYPE and text in c.get('text', ''):
if type == 'image':
content_node = {
'type': 'image',
'img_path': element.get('image_path'),
'img_alt': '',
'img_title': '',
'img_caption': '',
}
elif type == 'table':
content_node = {
'type': 'table',
'img_path': element.get('image_path'),
'table_latex': element.get('text'),
'table_title': '',
'table_caption': '',
'table_quality': element.get('quality'),
}
content_list.insert(i + 1, content_node)
break
else:
logger.error(
f"Can't find the location of image {element.get('image_path')} in the markdown file, search target is {text}"
)
def __insert_before_para(text, type, element, content_list):
"""在content_list中找到text，将image_path作为一个新的node插入到text前面."""
for i, c in enumerate(content_list):
content_type = c.get('type')
if content_type in UNI_FORMAT_TEXT_TYPE and text in c.get('text', ''):
if type == 'image':
content_node = {
'type': 'image',
'img_path': element.get('image_path'),
'img_alt': '',
'img_title': '',
'img_caption': '',
}
elif type == 'table':
content_node = {
'type': 'table',
'img_path': element.get('image_path'),
'table_latex': element.get('text'),
'table_title': '',
'table_caption': '',
'table_quality': element.get('quality'),
}
content_list.insert(i, content_node)
break
else:
logger.error(
f"Can't find the location of image {element.get('image_path')} in the markdown file, search target is {text}"
)
def mk_universal_format(pdf_info_list: list, img_buket_path):
"""构造统一格式 https://aicarrier.feishu.cn/wiki/FqmMwcH69iIdCWkkyjvcDwNUnTY."""
content_lst = []
for page_info in pdf_info_list:
page_lst = [] # 一个page内的段落列表
para_blocks = page_info.get('para_blocks')
pymu_raw_blocks = page_info.get('preproc_blocks')
all_page_images = []
all_page_images.extend(page_info.get('images', []))
all_page_images.extend(page_info.get('image_backup', []))
# all_page_images.extend(page_info.get("tables",[]))
# all_page_images.extend(page_info.get("table_backup",[]) )
all_page_tables = []
all_page_tables.extend(page_info.get('tables', []))
if not para_blocks or not pymu_raw_blocks: # 只有图片的拼接的场景
for img in all_page_images:
content_node = {
'type': 'image',
'img_path': join_path(img_buket_path, img['image_path']),
'img_alt': '',
'img_title': '',
'img_caption': '',
}
page_lst.append(content_node) # TODO 图片顺序
for table in all_page_tables:
content_node = {
'type': 'table',
'img_path': join_path(img_buket_path, table['image_path']),
'table_latex': table.get('text'),
'table_title': '',
'table_caption': '',
'table_quality': table.get('quality'),
}
page_lst.append(content_node) # TODO 图片顺序
else:
for block in para_blocks:
item = block['paras']
for _, p in item.items():
font_type = p[
'para_font_type'
] # 对于文本来说，要么是普通文本，要么是个行间公式
if font_type == TYPE_INTERLINE_EQUATION:
content_node = {'type': 'equation', 'latex': p['para_text']}
page_lst.append(content_node)
else:
para_text = p['para_text']
is_title = p['is_para_title']
title_level = p['para_title_level']
if is_title:
content_node = {
'type': f'h{title_level}',
'text': para_text,
}
page_lst.append(content_node)
else:
content_node = {'type': 'text', 'text': para_text}
page_lst.append(content_node)
content_lst.extend(page_lst)
"""插入图片"""
for img in all_page_images:
insert_img_or_table('image', img, pymu_raw_blocks, content_lst)
"""插入表格"""
for table in all_page_tables:
insert_img_or_table('table', table, pymu_raw_blocks, content_lst)
# end for
return content_lst
def insert_img_or_table(type, element, pymu_raw_blocks, content_lst):
element_bbox = element['bbox']
# 先看在哪个block内
for block in pymu_raw_blocks:
bbox = block['bbox']
if (
bbox[0] - 1 <= element_bbox[0] < bbox[2] + 1
and bbox[1] - 1 <= element_bbox[1] < bbox[3] + 1
): # 确定在这个大的block内，然后进入逐行比较距离
for l in block['lines']: # noqa: E741
line_box = l['bbox']
if (
line_box[0] - 1 <= element_bbox[0] < line_box[2] + 1
and line_box[1] - 1 <= element_bbox[1] < line_box[3] + 1
): # 在line内的，插入line前面
line_txt = ''.join([s['text'] for s in l['spans']])
__insert_before_para(line_txt, type, element, content_lst)
break
break
else: # 在行与行之间
# 找到图片x0,y0与line的x0,y0最近的line
min_distance = 100000
min_line = None
for l in block['lines']: # noqa: E741
line_box = l['bbox']
distance = math.sqrt(
(line_box[0] - element_bbox[0]) ** 2
+ (line_box[1] - element_bbox[1]) ** 2
)
if distance < min_distance:
min_distance = distance
min_line = l
if min_line:
line_txt = ''.join([s['text'] for s in min_line['spans']])
img_h = element_bbox[3] - element_bbox[1]
if min_distance < img_h: # 文字在图片前面
__insert_after_para(line_txt, type, element, content_lst)
else:
__insert_before_para(line_txt, type, element, content_lst)
break
else:
logger.error(
f"Can't find the location of image {element.get('image_path')} in the markdown file #1"
)
else: # 应当在两个block之间
# 找到上方最近的block，如果上方没有就找大下方最近的block
top_txt_block = find_top_nearest_text_bbox(pymu_raw_blocks, element_bbox)
if top_txt_block:
line_txt = ''.join([s['text'] for s in top_txt_block['lines'][-1]['spans']])
__insert_after_para(line_txt, type, element, content_lst)
else:
bottom_txt_block = find_bottom_nearest_text_bbox(
pymu_raw_blocks, element_bbox
)
if bottom_txt_block:
line_txt = ''.join(
[s['text'] for s in bottom_txt_block['lines'][0]['spans']]
)
__insert_before_para(line_txt, type, element, content_lst)
else: # TODO ，图片可能独占一列，这种情况上下是没有图片的
logger.error(
f"Can't find the location of image {element.get('image_path')} in the markdown file #2"
)
def mk_mm_markdown(content_list):
"""基于同一格式的内容列表，构造markdown，含图片."""
content_md = []
for c in content_list:
content_type = c.get('type')
if content_type == 'text':
content_md.append(c.get('text'))
elif content_type == 'equation':
content = c.get('latex')
if content.startswith('$$') and content.endswith('$$'):
content_md.append(content)
else:
content_md.append(f"\n$$\n{c.get('latex')}\n$$\n")
elif content_type in UNI_FORMAT_TEXT_TYPE:
content_md.append(f"{'#'*int(content_type[1])} {c.get('text')}")
elif content_type == 'image':
content_md.append(f"![]({c.get('img_path')})")
return '\n\n'.join(content_md)
def mk_nlp_markdown(content_list):
"""基于同一格式的内容列表，构造markdown，不含图片."""
content_md = []
for c in content_list:
content_type = c.get('type')
if content_type == 'text':
content_md.append(c.get('text'))
elif content_type == 'equation':
content_md.append(f"$$\n{c.get('latex')}\n$$")
elif content_type == 'table':
content_md.append(f"$$$\n{c.get('table_latex')}\n$$$")
elif content_type in UNI_FORMAT_TEXT_TYPE:
content_md.append(f"{'#'*int(content_type[1])} {c.get('text')}")
return '\n\n'.join(content_md)
magic_pdf/filter/pdf_meta_scan.py
View file @ 8afff9ae
"""输入： s3路径，每行一个 输出： pdf文件元信息，包括每一页上的所有图片的长宽高，bbox位置.""" """输入： s3路径，每行一个 输出： pdf文件元信息，包括每一页上的所有图片的长宽高，bbox位置."""
import sys
from collections import Counter from collections import Counter
import click import fitz
from loguru import logger from loguru import logger
from magic_pdf.config.drop_reason import DropReason from magic_pdf.config.drop_reason import DropReason
from magic_pdf.libs.commons import fitz, get_top_percent_list, mymax, read_file from magic_pdf.libs.commons import get_top_percent_list, mymax
from magic_pdf.libs.language import detect_lang from magic_pdf.libs.language import detect_lang
from magic_pdf.libs.pdf_check import detect_invalid_chars from magic_pdf.libs.pdf_check import detect_invalid_chars
...@@ -384,21 +383,8 @@ def pdf_meta_scan(pdf_bytes: bytes): ...@@ -384,21 +383,8 @@ def pdf_meta_scan(pdf_bytes: bytes):
return res return res
@click.command()
@click.option('--s3-pdf-path', help='s3上pdf文件的路径')
@click.option('--s3-profile', help='s3上的profile')
def main(s3_pdf_path: str, s3_profile: str):
""""""
try:
file_content = read_file(s3_pdf_path, s3_profile)
pdf_meta_scan(file_content)
except Exception as e:
print(f'ERROR: {s3_pdf_path}, {e}', file=sys.stderr)
logger.exception(e)
if __name__ == '__main__': if __name__ == '__main__':
main() pass
# "D:\project/20231108code-clean\pdf_cost_time\竖排例子\净空法师-大乘无量寿.pdf" # "D:\project/20231108code-clean\pdf_cost_time\竖排例子\净空法师-大乘无量寿.pdf"
# "D:\project/20231108code-clean\pdf_cost_time\竖排例子\三国演义_繁体竖排版.pdf" # "D:\project/20231108code-clean\pdf_cost_time\竖排例子\三国演义_繁体竖排版.pdf"
# "D:\project/20231108code-clean\pdf_cost_time\scihub\scihub_86800000\libgen.scimag86880000-86880999.zip_10.1021/acsami.1c03109.s002.pdf" # "D:\project/20231108code-clean\pdf_cost_time\scihub\scihub_86800000\libgen.scimag86880000-86880999.zip_10.1021/acsami.1c03109.s002.pdf"
......
magic_pdf/layout/bbox_sort.py deleted 100644 → 0
View file @ 4df1eb74
# 定义这里的bbox是一个list [x0, y0, x1, y1, block_content, idx_x, idx_y, content_type, ext_x0, ext_y0, ext_x1, ext_y1], 初始时候idx_x, idx_y都是None
# 其中x0, y0代表左上角坐标，x1, y1代表右下角坐标，坐标原点在左上角。
from magic_pdf.layout.layout_spiler_recog import get_spilter_of_page
from magic_pdf.libs.boxbase import _is_in, _is_in_or_part_overlap, _is_vertical_full_overlap
from magic_pdf.libs.commons import mymax
X0_IDX = 0
Y0_IDX = 1
X1_IDX = 2
Y1_IDX = 3
CONTENT_IDX = 4
IDX_X = 5
IDX_Y = 6
CONTENT_TYPE_IDX = 7
X0_EXT_IDX = 8
Y0_EXT_IDX = 9
X1_EXT_IDX = 10
Y1_EXT_IDX = 11
def prepare_bboxes_for_layout_split(image_info, image_backup_info, table_info, inline_eq_info, interline_eq_info, text_raw_blocks: dict, page_boundry, page):
"""
text_raw_blocks:结构参考test/assets/papre/pymu_textblocks.json
把bbox重新组装成一个list，每个元素[x0, y0, x1, y1, block_content, idx_x, idx_y, content_type, ext_x0, ext_y0, ext_x1, ext_y1], 初始时候idx_x, idx_y都是None. 对于图片、公式来说，block_content是图片的地址， 对于段落来说，block_content是pymupdf里的block结构
"""
all_bboxes = []
for image in image_info:
box = image['bbox']
# 由于没有实现横向的栏切分，因此在这里先过滤掉一些小的图片。这些图片有可能影响layout，造成没有横向栏切分的情况下，layout切分不准确。例如 scihub_76500000/libgen.scimag76570000-76570999.zip_10.1186/s13287-019-1355-1
# 把长宽都小于50的去掉
if abs(box[0]-box[2]) < 50 and abs(box[1]-box[3]) < 50:
continue
all_bboxes.append([box[0], box[1], box[2], box[3], None, None, None, 'image', None, None, None, None])
for table in table_info:
box = table['bbox']
all_bboxes.append([box[0], box[1], box[2], box[3], None, None, None, 'table', None, None, None, None])
"""由于公式与段落混合，因此公式不再参与layout划分，无需加入all_bboxes"""
# 加入文本block
text_block_temp = []
for block in text_raw_blocks:
bbox = block['bbox']
text_block_temp.append([bbox[0], bbox[1], bbox[2], bbox[3], None, None, None, 'text', None, None, None, None])
text_block_new = resolve_bbox_overlap_for_layout_det(text_block_temp)
text_block_new = filter_lines_bbox(text_block_new) # 去掉线条bbox，有可能让layout探测陷入无限循环
"""找出会影响layout的色块、横向分割线"""
spilter_bboxes = get_spilter_of_page(page, [b['bbox'] for b in image_info]+[b['bbox'] for b in image_backup_info], [b['bbox'] for b in table_info], )
# 还要去掉存在于spilter_bboxes里的text_block
if len(spilter_bboxes) > 0:
text_block_new = [box for box in text_block_new if not any([_is_in_or_part_overlap(box[:4], spilter_bbox) for spilter_bbox in spilter_bboxes])]
for bbox in text_block_new:
all_bboxes.append([bbox[0], bbox[1], bbox[2], bbox[3], None, None, None, 'text', None, None, None, None])
for bbox in spilter_bboxes:
all_bboxes.append([bbox[0], bbox[1], bbox[2], bbox[3], None, None, None, 'spilter', None, None, None, None])
return all_bboxes
def resolve_bbox_overlap_for_layout_det(bboxes:list):
"""
1. 去掉bbox互相包含的，去掉被包含的
2. 上下方向上如果有重叠，就扩大大box范围，直到覆盖小box
"""
def _is_in_other_bbox(i:int):
"""
判断i个box是否被其他box有所包含
"""
for j in range(0, len(bboxes)):
if j!=i and _is_in(bboxes[i][:4], bboxes[j][:4]):
return True
# elif j!=i and _is_bottom_full_overlap(bboxes[i][:4], bboxes[j][:4]):
# return True
return False
# 首先去掉被包含的bbox
new_bbox_1 = []
for i in range(0, len(bboxes)):
if not _is_in_other_bbox(i):
new_bbox_1.append(bboxes[i])
# 其次扩展大的box
new_box = []
new_bbox_2 = []
len_1 = len(new_bbox_2)
while True:
merged_idx = []
for i in range(0, len(new_bbox_1)):
if i in merged_idx:
continue
for j in range(i+1, len(new_bbox_1)):
if j in merged_idx:
continue
bx1 = new_bbox_1[i]
bx2 = new_bbox_1[j]
if i!=j and _is_vertical_full_overlap(bx1[:4], bx2[:4]):
merged_box = min([bx1[0], bx2[0]]), min([bx1[1], bx2[1]]), max([bx1[2], bx2[2]]), max([bx1[3], bx2[3]])
new_bbox_2.append(merged_box)
merged_idx.append(i)
merged_idx.append(j)
for i in range(0, len(new_bbox_1)): # 没有合并的加入进来
if i not in merged_idx:
new_bbox_2.append(new_bbox_1[i])
if len(new_bbox_2)==0 or len_1==len(new_bbox_2):
break
else:
len_1 = len(new_bbox_2)
new_box = new_bbox_2
new_bbox_1, new_bbox_2 = new_bbox_2, []
return new_box
def filter_lines_bbox(bboxes: list):
"""
过滤掉bbox为空的行
"""
new_box = []
for box in bboxes:
x0, y0, x1, y1 = box[0], box[1], box[2], box[3]
if abs(x0-x1)<=1 or abs(y0-y1)<=1:
continue
else:
new_box.append(box)
return new_box
################################################################################
# 第一种排序算法
# 以下是基于延长线遮挡做的一个算法
#
################################################################################
def find_all_left_bbox(this_bbox, all_bboxes) -> list:
"""
寻找this_bbox左边的所有bbox
"""
left_boxes = [box for box in all_bboxes if box[X1_IDX] <= this_bbox[X0_IDX]]
return left_boxes
def find_all_top_bbox(this_bbox, all_bboxes) -> list:
"""
寻找this_bbox上面的所有bbox
"""
top_boxes = [box for box in all_bboxes if box[Y1_IDX] <= this_bbox[Y0_IDX]]
return top_boxes
def get_and_set_idx_x(this_bbox, all_bboxes) -> int:
"""
寻找this_bbox在all_bboxes中的遮挡深度 idx_x
"""
if this_bbox[IDX_X] is not None:
return this_bbox[IDX_X]
else:
all_left_bboxes = find_all_left_bbox(this_bbox, all_bboxes)
if len(all_left_bboxes) == 0:
this_bbox[IDX_X] = 0
else:
all_left_bboxes_idx = [get_and_set_idx_x(bbox, all_bboxes) for bbox in all_left_bboxes]
max_idx_x = mymax(all_left_bboxes_idx)
this_bbox[IDX_X] = max_idx_x + 1
return this_bbox[IDX_X]
def get_and_set_idx_y(this_bbox, all_bboxes) -> int:
"""
寻找this_bbox在all_bboxes中y方向的遮挡深度 idx_y
"""
if this_bbox[IDX_Y] is not None:
return this_bbox[IDX_Y]
else:
all_top_bboxes = find_all_top_bbox(this_bbox, all_bboxes)
if len(all_top_bboxes) == 0:
this_bbox[IDX_Y] = 0
else:
all_top_bboxes_idx = [get_and_set_idx_y(bbox, all_bboxes) for bbox in all_top_bboxes]
max_idx_y = mymax(all_top_bboxes_idx)
this_bbox[IDX_Y] = max_idx_y + 1
return this_bbox[IDX_Y]
def bbox_sort(all_bboxes: list):
"""
排序
"""
all_bboxes_idx_x = [get_and_set_idx_x(bbox, all_bboxes) for bbox in all_bboxes]
all_bboxes_idx_y = [get_and_set_idx_y(bbox, all_bboxes) for bbox in all_bboxes]
all_bboxes_idx = [(idx_x, idx_y) for idx_x, idx_y in zip(all_bboxes_idx_x, all_bboxes_idx_y)]
all_bboxes_idx = [idx_x_y[0] * 100000 + idx_x_y[1] for idx_x_y in all_bboxes_idx] # 变换成一个点，保证能够先X，X相同时按Y排序
all_bboxes_idx = list(zip(all_bboxes_idx, all_bboxes))
all_bboxes_idx.sort(key=lambda x: x[0])
sorted_bboxes = [bbox for idx, bbox in all_bboxes_idx]
return sorted_bboxes
################################################################################
# 第二种排序算法
# 下面的算法在计算idx_x和idx_y的时候不考虑延长线，而只考虑实际的长或者宽被遮挡的情况
#
################################################################################
def find_left_nearest_bbox(this_bbox, all_bboxes) -> list:
"""
在all_bboxes里找到所有右侧高度和this_bbox有重叠的bbox
"""
left_boxes = [box for box in all_bboxes if box[X1_IDX] <= this_bbox[X0_IDX] and any([
box[Y0_IDX] < this_bbox[Y0_IDX] < box[Y1_IDX], box[Y0_IDX] < this_bbox[Y1_IDX] < box[Y1_IDX],
this_bbox[Y0_IDX] < box[Y0_IDX] < this_bbox[Y1_IDX], this_bbox[Y0_IDX] < box[Y1_IDX] < this_bbox[Y1_IDX],
box[Y0_IDX]==this_bbox[Y0_IDX] and box[Y1_IDX]==this_bbox[Y1_IDX]])]
# 然后再过滤一下，找到水平上距离this_bbox最近的那个
if len(left_boxes) > 0:
left_boxes.sort(key=lambda x: x[X1_IDX], reverse=True)
left_boxes = [left_boxes[0]]
else:
left_boxes = []
return left_boxes
def get_and_set_idx_x_2(this_bbox, all_bboxes):
"""
寻找this_bbox在all_bboxes中的被直接遮挡的深度 idx_x
这个遮挡深度不考虑延长线，而是被实际的长或者宽遮挡的情况
"""
if this_bbox[IDX_X] is not None:
return this_bbox[IDX_X]
else:
left_nearest_bbox = find_left_nearest_bbox(this_bbox, all_bboxes)
if len(left_nearest_bbox) == 0:
this_bbox[IDX_X] = 0
else:
left_idx_x = get_and_set_idx_x_2(left_nearest_bbox[0], all_bboxes)
this_bbox[IDX_X] = left_idx_x + 1
return this_bbox[IDX_X]
def find_top_nearest_bbox(this_bbox, all_bboxes) -> list:
"""
在all_bboxes里找到所有下侧宽度和this_bbox有重叠的bbox
"""
top_boxes = [box for box in all_bboxes if box[Y1_IDX] <= this_bbox[Y0_IDX] and any([
box[X0_IDX] < this_bbox[X0_IDX] < box[X1_IDX], box[X0_IDX] < this_bbox[X1_IDX] < box[X1_IDX],
this_bbox[X0_IDX] < box[X0_IDX] < this_bbox[X1_IDX], this_bbox[X0_IDX] < box[X1_IDX] < this_bbox[X1_IDX],
box[X0_IDX]==this_bbox[X0_IDX] and box[X1_IDX]==this_bbox[X1_IDX]])]
# 然后再过滤一下，找到水平上距离this_bbox最近的那个
if len(top_boxes) > 0:
top_boxes.sort(key=lambda x: x[Y1_IDX], reverse=True)
top_boxes = [top_boxes[0]]
else:
top_boxes = []
return top_boxes
def get_and_set_idx_y_2(this_bbox, all_bboxes):
"""
寻找this_bbox在all_bboxes中的被直接遮挡的深度 idx_y
这个遮挡深度不考虑延长线，而是被实际的长或者宽遮挡的情况
"""
if this_bbox[IDX_Y] is not None:
return this_bbox[IDX_Y]
else:
top_nearest_bbox = find_top_nearest_bbox(this_bbox, all_bboxes)
if len(top_nearest_bbox) == 0:
this_bbox[IDX_Y] = 0
else:
top_idx_y = get_and_set_idx_y_2(top_nearest_bbox[0], all_bboxes)
this_bbox[IDX_Y] = top_idx_y + 1
return this_bbox[IDX_Y]
def paper_bbox_sort(all_bboxes: list, page_width, page_height):
all_bboxes_idx_x = [get_and_set_idx_x_2(bbox, all_bboxes) for bbox in all_bboxes]
all_bboxes_idx_y = [get_and_set_idx_y_2(bbox, all_bboxes) for bbox in all_bboxes]
all_bboxes_idx = [(idx_x, idx_y) for idx_x, idx_y in zip(all_bboxes_idx_x, all_bboxes_idx_y)]
all_bboxes_idx = [idx_x_y[0] * 100000 + idx_x_y[1] for idx_x_y in all_bboxes_idx] # 变换成一个点，保证能够先X，X相同时按Y排序
all_bboxes_idx = list(zip(all_bboxes_idx, all_bboxes))
all_bboxes_idx.sort(key=lambda x: x[0])
sorted_bboxes = [bbox for idx, bbox in all_bboxes_idx]
return sorted_bboxes
################################################################################
"""
第三种排序算法, 假设page的最左侧为X0，最右侧为X1，最上侧为Y0，最下侧为Y1
这个排序算法在第二种算法基础上增加对bbox的预处理步骤。预处理思路如下：
1. 首先在水平方向上对bbox进行扩展。扩展方法是：
- 对每个bbox，找到其左边最近的bbox（也就是y方向有重叠），然后将其左边界扩展到左边最近bbox的右边界(x1+1),这里加1是为了避免重叠。如果没有左边的bbox，那么就将其左边界扩展到page的最左侧X0。
- 对每个bbox，找到其右边最近的bbox（也就是y方向有重叠），然后将其右边界扩展到右边最近bbox的左边界(x0-1),这里减1是为了避免重叠。如果没有右边的bbox，那么就将其右边界扩展到page的最右侧X1。
- 经过上面2个步骤，bbox扩展到了水平方向的最大范围。[左最近bbox.x1+1, 右最近bbox.x0-1]
2. 合并所有的连续水平方向的bbox, 合并方法是：
- 对bbox进行y方向排序，然后从上到下遍历所有bbox，如果当前bbox和下一个bbox的x0, x1等于X0, X1，那么就合并这两个bbox。
3. 然后在垂直方向上对bbox进行扩展。扩展方法是：
- 首先从page上切割掉合并后的水平bbox, 得到几个新的block
针对每个block
- x0: 扎到位于左侧x=x0延长线的左侧所有的bboxes, 找到最大的x1,让x0=x1+1。如果没有，则x0=X0
- x1: 找到位于右侧x=x1延长线右侧所有的bboxes， 找到最小的x0, 让x1=x0-1。如果没有，则x1=X1
随后在垂直方向上合并所有的连续的block，方法如下：
- 对block进行x方向排序，然后从左到右遍历所有block，如果当前block和下一个block的x0, x1相等，那么就合并这两个block。
如果垂直切分后所有小bbox都被分配到了一个block, 那么分割就完成了。这些合并后的block打上标签'GOOD_LAYOUT’
如果在某个垂直方向上无法被完全分割到一个block，那么就将这个block打上标签'BAD_LAYOUT'。
至此完成，一个页面的预处理，天然的block要么属于'GOOD_LAYOUT'，要么属于'BAD_LAYOUT'。针对含有'BAD_LAYOUT'的页面，可以先按照自上而下，自左到右进行天然排序，也可以先过滤掉这种书籍。
(完成条件下次加强：进行水平方向切分，把混乱的layout部分尽可能切割出去)
"""
################################################################################
def find_left_neighbor_bboxes(this_bbox, all_bboxes) -> list:
"""
在all_bboxes里找到所有右侧高度和this_bbox有重叠的bbox
这里使用扩展之后的bbox
"""
left_boxes = [box for box in all_bboxes if box[X1_EXT_IDX] <= this_bbox[X0_EXT_IDX] and any([
box[Y0_EXT_IDX] < this_bbox[Y0_EXT_IDX] < box[Y1_EXT_IDX], box[Y0_EXT_IDX] < this_bbox[Y1_EXT_IDX] < box[Y1_EXT_IDX],
this_bbox[Y0_EXT_IDX] < box[Y0_EXT_IDX] < this_bbox[Y1_EXT_IDX], this_bbox[Y0_EXT_IDX] < box[Y1_EXT_IDX] < this_bbox[Y1_EXT_IDX],
box[Y0_EXT_IDX]==this_bbox[Y0_EXT_IDX] and box[Y1_EXT_IDX]==this_bbox[Y1_EXT_IDX]])]
# 然后再过滤一下，找到水平上距离this_bbox最近的那个
if len(left_boxes) > 0:
left_boxes.sort(key=lambda x: x[X1_EXT_IDX], reverse=True)
left_boxes = left_boxes
else:
left_boxes = []
return left_boxes
def find_top_neighbor_bboxes(this_bbox, all_bboxes) -> list:
"""
在all_bboxes里找到所有下侧宽度和this_bbox有重叠的bbox
这里使用扩展之后的bbox
"""
top_boxes = [box for box in all_bboxes if box[Y1_EXT_IDX] <= this_bbox[Y0_EXT_IDX] and any([
box[X0_EXT_IDX] < this_bbox[X0_EXT_IDX] < box[X1_EXT_IDX], box[X0_EXT_IDX] < this_bbox[X1_EXT_IDX] < box[X1_EXT_IDX],
this_bbox[X0_EXT_IDX] < box[X0_EXT_IDX] < this_bbox[X1_EXT_IDX], this_bbox[X0_EXT_IDX] < box[X1_EXT_IDX] < this_bbox[X1_EXT_IDX],
box[X0_EXT_IDX]==this_bbox[X0_EXT_IDX] and box[X1_EXT_IDX]==this_bbox[X1_EXT_IDX]])]
# 然后再过滤一下，找到水平上距离this_bbox最近的那个
if len(top_boxes) > 0:
top_boxes.sort(key=lambda x: x[Y1_EXT_IDX], reverse=True)
top_boxes = top_boxes
else:
top_boxes = []
return top_boxes
def get_and_set_idx_x_2_ext(this_bbox, all_bboxes):
"""
寻找this_bbox在all_bboxes中的被直接遮挡的深度 idx_x
这个遮挡深度不考虑延长线，而是被实际的长或者宽遮挡的情况
"""
if this_bbox[IDX_X] is not None:
return this_bbox[IDX_X]
else:
left_nearest_bbox = find_left_neighbor_bboxes(this_bbox, all_bboxes)
if len(left_nearest_bbox) == 0:
this_bbox[IDX_X] = 0
else:
left_idx_x = [get_and_set_idx_x_2(b, all_bboxes) for b in left_nearest_bbox]
this_bbox[IDX_X] = mymax(left_idx_x) + 1
return this_bbox[IDX_X]
def get_and_set_idx_y_2_ext(this_bbox, all_bboxes):
"""
寻找this_bbox在all_bboxes中的被直接遮挡的深度 idx_y
这个遮挡深度不考虑延长线，而是被实际的长或者宽遮挡的情况
"""
if this_bbox[IDX_Y] is not None:
return this_bbox[IDX_Y]
else:
top_nearest_bbox = find_top_neighbor_bboxes(this_bbox, all_bboxes)
if len(top_nearest_bbox) == 0:
this_bbox[IDX_Y] = 0
else:
top_idx_y = [get_and_set_idx_y_2_ext(b, all_bboxes) for b in top_nearest_bbox]
this_bbox[IDX_Y] = mymax(top_idx_y) + 1
return this_bbox[IDX_Y]
def _paper_bbox_sort_ext(all_bboxes: list):
all_bboxes_idx_x = [get_and_set_idx_x_2_ext(bbox, all_bboxes) for bbox in all_bboxes]
all_bboxes_idx_y = [get_and_set_idx_y_2_ext(bbox, all_bboxes) for bbox in all_bboxes]
all_bboxes_idx = [(idx_x, idx_y) for idx_x, idx_y in zip(all_bboxes_idx_x, all_bboxes_idx_y)]
all_bboxes_idx = [idx_x_y[0] * 100000 + idx_x_y[1] for idx_x_y in all_bboxes_idx] # 变换成一个点，保证能够先X，X相同时按Y排序
all_bboxes_idx = list(zip(all_bboxes_idx, all_bboxes))
all_bboxes_idx.sort(key=lambda x: x[0])
sorted_bboxes = [bbox for idx, bbox in all_bboxes_idx]
return sorted_bboxes
# ===============================================================================================
def find_left_bbox_ext_line(this_bbox, all_bboxes) -> list:
"""
寻找this_bbox左边的所有bbox, 使用延长线
"""
left_boxes = [box for box in all_bboxes if box[X1_IDX] <= this_bbox[X0_IDX]]
if len(left_boxes):
left_boxes.sort(key=lambda x: x[X1_IDX], reverse=True)
left_boxes = left_boxes[0]
else:
left_boxes = None
return left_boxes
def find_right_bbox_ext_line(this_bbox, all_bboxes) -> list:
"""
寻找this_bbox右边的所有bbox, 使用延长线
"""
right_boxes = [box for box in all_bboxes if box[X0_IDX] >= this_bbox[X1_IDX]]
if len(right_boxes):
right_boxes.sort(key=lambda x: x[X0_IDX])
right_boxes = right_boxes[0]
else:
right_boxes = None
return right_boxes
# =============================================================================================
def find_left_nearest_bbox_direct(this_bbox, all_bboxes) -> list:
"""
在all_bboxes里找到所有右侧高度和this_bbox有重叠的bbox， 不用延长线并且不能像
"""
left_boxes = [box for box in all_bboxes if box[X1_IDX] <= this_bbox[X0_IDX] and any([
box[Y0_IDX] < this_bbox[Y0_IDX] < box[Y1_IDX], box[Y0_IDX] < this_bbox[Y1_IDX] < box[Y1_IDX],
this_bbox[Y0_IDX] < box[Y0_IDX] < this_bbox[Y1_IDX], this_bbox[Y0_IDX] < box[Y1_IDX] < this_bbox[Y1_IDX],
box[Y0_IDX]==this_bbox[Y0_IDX] and box[Y1_IDX]==this_bbox[Y1_IDX]])]
# 然后再过滤一下，找到水平上距离this_bbox最近的那个——x1最大的那个
if len(left_boxes) > 0:
left_boxes.sort(key=lambda x: x[X1_EXT_IDX] if x[X1_EXT_IDX] else x[X1_IDX], reverse=True)
left_boxes = left_boxes[0]
else:
left_boxes = None
return left_boxes
def find_right_nearst_bbox_direct(this_bbox, all_bboxes) -> list:
"""
找到在this_bbox右侧且距离this_bbox距离最近的bbox.必须是直接遮挡的那种
"""
right_bboxes = [box for box in all_bboxes if box[X0_IDX] >= this_bbox[X1_IDX] and any([
this_bbox[Y0_IDX] < box[Y0_IDX] < this_bbox[Y1_IDX], this_bbox[Y0_IDX] < box[Y1_IDX] < this_bbox[Y1_IDX],
box[Y0_IDX] < this_bbox[Y0_IDX] < box[Y1_IDX], box[Y0_IDX] < this_bbox[Y1_IDX] < box[Y1_IDX],
box[Y0_IDX]==this_bbox[Y0_IDX] and box[Y1_IDX]==this_bbox[Y1_IDX]])]
if len(right_bboxes)>0:
right_bboxes.sort(key=lambda x: x[X0_EXT_IDX] if x[X0_EXT_IDX] else x[X0_IDX])
right_bboxes = right_bboxes[0]
else:
right_bboxes = None
return right_bboxes
def reset_idx_x_y(all_boxes:list)->list:
for box in all_boxes:
box[IDX_X] = None
box[IDX_Y] = None
return all_boxes
# ===================================================================================================
def find_top_nearest_bbox_direct(this_bbox, bboxes_collection) -> list:
"""
找到在this_bbox上方且距离this_bbox距离最近的bbox.必须是直接遮挡的那种
"""
top_bboxes = [box for box in bboxes_collection if box[Y1_IDX] <= this_bbox[Y0_IDX] and any([
box[X0_IDX] < this_bbox[X0_IDX] < box[X1_IDX], box[X0_IDX] < this_bbox[X1_IDX] < box[X1_IDX],
this_bbox[X0_IDX] < box[X0_IDX] < this_bbox[X1_IDX], this_bbox[X0_IDX] < box[X1_IDX] < this_bbox[X1_IDX],
box[X0_IDX]==this_bbox[X0_IDX] and box[X1_IDX]==this_bbox[X1_IDX]])]
# 然后再过滤一下，找到上方距离this_bbox最近的那个
if len(top_bboxes) > 0:
top_bboxes.sort(key=lambda x: x[Y1_IDX], reverse=True)
top_bboxes = top_bboxes[0]
else:
top_bboxes = None
return top_bboxes
def find_bottom_nearest_bbox_direct(this_bbox, bboxes_collection) -> list:
"""
找到在this_bbox下方且距离this_bbox距离最近的bbox.必须是直接遮挡的那种
"""
bottom_bboxes = [box for box in bboxes_collection if box[Y0_IDX] >= this_bbox[Y1_IDX] and any([
box[X0_IDX] < this_bbox[X0_IDX] < box[X1_IDX], box[X0_IDX] < this_bbox[X1_IDX] < box[X1_IDX],
this_bbox[X0_IDX] < box[X0_IDX] < this_bbox[X1_IDX], this_bbox[X0_IDX] < box[X1_IDX] < this_bbox[X1_IDX],
box[X0_IDX]==this_bbox[X0_IDX] and box[X1_IDX]==this_bbox[X1_IDX]])]
# 然后再过滤一下，找到水平上距离this_bbox最近的那个
if len(bottom_bboxes) > 0:
bottom_bboxes.sort(key=lambda x: x[Y0_IDX])
bottom_bboxes = bottom_bboxes[0]
else:
bottom_bboxes = None
return bottom_bboxes
def find_boundry_bboxes(bboxes:list) -> tuple:
"""
找到bboxes的边界——找到所有bbox里最小的(x0, y0), 最大的(x1, y1)
"""
x0, y0, x1, y1 = bboxes[0][X0_IDX], bboxes[0][Y0_IDX], bboxes[0][X1_IDX], bboxes[0][Y1_IDX]
for box in bboxes:
x0 = min(box[X0_IDX], x0)
y0 = min(box[Y0_IDX], y0)
x1 = max(box[X1_IDX], x1)
y1 = max(box[Y1_IDX], y1)
return x0, y0, x1, y1
def extend_bbox_vertical(bboxes:list, boundry_x0, boundry_y0, boundry_x1, boundry_y1) -> list:
"""
在垂直方向上扩展能够直接垂直打通的bbox,也就是那些上下都没有其他box的bbox
"""
for box in bboxes:
top_nearest_bbox = find_top_nearest_bbox_direct(box, bboxes)
bottom_nearest_bbox = find_bottom_nearest_bbox_direct(box, bboxes)
if top_nearest_bbox is None and bottom_nearest_bbox is None: # 独占一列
box[X0_EXT_IDX] = box[X0_IDX]
box[Y0_EXT_IDX] = boundry_y0
box[X1_EXT_IDX] = box[X1_IDX]
box[Y1_EXT_IDX] = boundry_y1
# else:
# if top_nearest_bbox is None:
# box[Y0_EXT_IDX] = boundry_y0
# else:
# box[Y0_EXT_IDX] = top_nearest_bbox[Y1_IDX] + 1
# if bottom_nearest_bbox is None:
# box[Y1_EXT_IDX] = boundry_y1
# else:
# box[Y1_EXT_IDX] = bottom_nearest_bbox[Y0_IDX] - 1
# box[X0_EXT_IDX] = box[X0_IDX]
# box[X1_EXT_IDX] = box[X1_IDX]
return bboxes
# ===================================================================================================
def paper_bbox_sort_v2(all_bboxes: list, page_width:int, page_height:int):
"""
增加预处理行为的排序:
return:
[
{
"layout_bbox": [x0, y0, x1, y1],
"layout_label":"GOOD_LAYOUT/BAD_LAYOUT",
"content_bboxes": [] #每个元素都是[x0, y0, x1, y1, block_content, idx_x, idx_y, content_type, ext_x0, ext_y0, ext_x1, ext_y1], 并且顺序就是阅读顺序
}
]
"""
sorted_layouts = [] # 最后的返回结果
page_x0, page_y0, page_x1, page_y1 = 1, 1, page_width-1, page_height-1
all_bboxes = paper_bbox_sort(all_bboxes) # 大致拍下序
# 首先在水平方向上扩展独占一行的bbox
for bbox in all_bboxes:
left_nearest_bbox = find_left_nearest_bbox_direct(bbox, all_bboxes) # 非扩展线
right_nearest_bbox = find_right_nearst_bbox_direct(bbox, all_bboxes)
if left_nearest_bbox is None and right_nearest_bbox is None: # 独占一行
bbox[X0_EXT_IDX] = page_x0
bbox[Y0_EXT_IDX] = bbox[Y0_IDX]
bbox[X1_EXT_IDX] = page_x1
bbox[Y1_EXT_IDX] = bbox[Y1_IDX]
# 此时独占一行的被成功扩展到指定的边界上，这个时候利用边界条件合并连续的bbox，成为一个group
if len(all_bboxes)==1:
return [{"layout_bbox": [page_x0, page_y0, page_x1, page_y1], "layout_label":"GOOD_LAYOUT", "content_bboxes": all_bboxes}]
if len(all_bboxes)==0:
return []
"""
然后合并所有连续水平方向的bbox.
"""
all_bboxes.sort(key=lambda x: x[Y0_IDX])
h_bboxes = []
h_bbox_group = []
v_boxes = []
for bbox in all_bboxes:
if bbox[X0_IDX] == page_x0 and bbox[X1_IDX] == page_x1:
h_bbox_group.append(bbox)
else:
if len(h_bbox_group)>0:
h_bboxes.append(h_bbox_group)
h_bbox_group = []
# 最后一个group
if len(h_bbox_group)>0:
h_bboxes.append(h_bbox_group)
"""
现在h_bboxes里面是所有的group了，每个group都是一个list
对h_bboxes里的每个group进行计算放回到sorted_layouts里
"""
for gp in h_bboxes:
gp.sort(key=lambda x: x[Y0_IDX])
block_info = {"layout_label":"GOOD_LAYOUT", "content_bboxes": gp}
# 然后计算这个group的layout_bbox，也就是最小的x0,y0, 最大的x1,y1
x0, y0, x1, y1 = gp[0][X0_EXT_IDX], gp[0][Y0_EXT_IDX], gp[-1][X1_EXT_IDX], gp[-1][Y1_EXT_IDX]
block_info["layout_bbox"] = [x0, y0, x1, y1]
sorted_layouts.append(block_info)
# 接下来利用这些连续的水平bbox的layout_bbox的y0, y1，从水平上切分开其余的为几个部分
h_split_lines = [page_y0]
for gp in h_bboxes:
layout_bbox = gp['layout_bbox']
y0, y1 = layout_bbox[1], layout_bbox[3]
h_split_lines.append(y0)
h_split_lines.append(y1)
h_split_lines.append(page_y1)
unsplited_bboxes = []
for i in range(0, len(h_split_lines), 2):
start_y0, start_y1 = h_split_lines[i:i+2]
# 然后找出[start_y0, start_y1]之间的其他bbox，这些组成一个未分割板块
bboxes_in_block = [bbox for bbox in all_bboxes if bbox[Y0_IDX]>=start_y0 and bbox[Y1_IDX]<=start_y1]
unsplited_bboxes.append(bboxes_in_block)
# ================== 至此，水平方向的 已经切分排序完毕====================================
"""
接下来针对每个非水平的部分切分垂直方向的
此时，只剩下了无法被完全水平打通的bbox了。对这些box，优先进行垂直扩展，然后进行垂直切分.
分3步：
1. 先把能完全垂直打通的隔离出去当做一个layout
2. 其余的先垂直切分
3. 垂直切分之后的部分再尝试水平切分
4. 剩下的不能被切分的各个部分当成一个layout
"""
# 对每部分进行垂直切分
for bboxes_in_block in unsplited_bboxes:
# 首先对这个block的bbox进行垂直方向上的扩展
boundry_x0, boundry_y0, boundry_x1, boundry_y1 = find_boundry_bboxes(bboxes_in_block)
# 进行垂直方向上的扩展
extended_vertical_bboxes = extend_bbox_vertical(bboxes_in_block, boundry_x0, boundry_y0, boundry_x1, boundry_y1)
# 然后对这个block进行垂直方向上的切分
extend_bbox_vertical.sort(key=lambda x: x[X0_IDX]) # x方向上从小到大，代表了从左到右读取
v_boxes_group = []
for bbox in extended_vertical_bboxes:
if bbox[Y0_IDX]==boundry_y0 and bbox[Y1_IDX]==boundry_y1:
v_boxes_group.append(bbox)
else:
if len(v_boxes_group)>0:
v_boxes.append(v_boxes_group)
v_boxes_group = []
if len(v_boxes_group)>0:
v_boxes.append(v_boxes_group)
# 把连续的垂直部分加入到sorted_layouts里。注意这个时候已经是连续的垂直部分了，因为上面已经做了
for gp in v_boxes:
gp.sort(key=lambda x: x[X0_IDX])
block_info = {"layout_label":"GOOD_LAYOUT", "content_bboxes": gp}
# 然后计算这个group的layout_bbox，也就是最小的x0,y0, 最大的x1,y1
x0, y0, x1, y1 = gp[0][X0_EXT_IDX], gp[0][Y0_EXT_IDX], gp[-1][X1_EXT_IDX], gp[-1][Y1_EXT_IDX]
block_info["layout_bbox"] = [x0, y0, x1, y1]
sorted_layouts.append(block_info)
# 在垂直方向上，划分子块，也就是用贯通的垂直线进行切分。这些被切分出来的块，极大可能是可被垂直切分的，如果不能完全的垂直切分，那么尝试水平切分。都不能的则当成一个layout
v_split_lines = [boundry_x0]
for gp in v_boxes:
layout_bbox = gp['layout_bbox']
x0, x1 = layout_bbox[0], layout_bbox[2]
v_split_lines.append(x0)
v_split_lines.append(x1)
v_split_lines.append(boundry_x1)
reset_idx_x_y(all_bboxes)
all_boxes = _paper_bbox_sort_ext(all_bboxes)
return all_boxes
magic_pdf/layout/layout_det_utils.py deleted 100644 → 0
View file @ 4df1eb74
from magic_pdf.layout.bbox_sort import X0_EXT_IDX, X0_IDX, X1_EXT_IDX, X1_IDX, Y0_IDX, Y1_EXT_IDX, Y1_IDX
from magic_pdf.libs.boxbase import _is_bottom_full_overlap, _left_intersect, _right_intersect
def find_all_left_bbox_direct(this_bbox, all_bboxes) -> list:
"""
在all_bboxes里找到所有右侧垂直方向上和this_bbox有重叠的bbox， 不用延长线
并且要考虑两个box左右相交的情况，如果相交了，那么右侧的box就不算最左侧。
"""
left_boxes = [box for box in all_bboxes if box[X1_IDX] <= this_bbox[X0_IDX]
and any([
box[Y0_IDX] < this_bbox[Y0_IDX] < box[Y1_IDX], box[Y0_IDX] < this_bbox[Y1_IDX] < box[Y1_IDX],
this_bbox[Y0_IDX] < box[Y0_IDX] < this_bbox[Y1_IDX], this_bbox[Y0_IDX] < box[Y1_IDX] < this_bbox[Y1_IDX],
box[Y0_IDX]==this_bbox[Y0_IDX] and box[Y1_IDX]==this_bbox[Y1_IDX]]) or _left_intersect(box[:4], this_bbox[:4])]
# 然后再过滤一下，找到水平上距离this_bbox最近的那个——x1最大的那个
if len(left_boxes) > 0:
left_boxes.sort(key=lambda x: x[X1_EXT_IDX] if x[X1_EXT_IDX] else x[X1_IDX], reverse=True)
left_boxes = left_boxes[0]
else:
left_boxes = None
return left_boxes
def find_all_right_bbox_direct(this_bbox, all_bboxes) -> list:
"""
找到在this_bbox右侧且距离this_bbox距离最近的bbox.必须是直接遮挡的那种
"""
right_bboxes = [box for box in all_bboxes if box[X0_IDX] >= this_bbox[X1_IDX]
and any([
this_bbox[Y0_IDX] < box[Y0_IDX] < this_bbox[Y1_IDX], this_bbox[Y0_IDX] < box[Y1_IDX] < this_bbox[Y1_IDX],
box[Y0_IDX] < this_bbox[Y0_IDX] < box[Y1_IDX], box[Y0_IDX] < this_bbox[Y1_IDX] < box[Y1_IDX],
box[Y0_IDX]==this_bbox[Y0_IDX] and box[Y1_IDX]==this_bbox[Y1_IDX]]) or _right_intersect(this_bbox[:4], box[:4])]
if len(right_bboxes)>0:
right_bboxes.sort(key=lambda x: x[X0_EXT_IDX] if x[X0_EXT_IDX] else x[X0_IDX])
right_bboxes = right_bboxes[0]
else:
right_bboxes = None
return right_bboxes
def find_all_top_bbox_direct(this_bbox, all_bboxes) -> list:
"""
找到在this_bbox上侧且距离this_bbox距离最近的bbox.必须是直接遮挡的那种
"""
top_bboxes = [box for box in all_bboxes if box[Y1_IDX] <= this_bbox[Y0_IDX] and any([
box[X0_IDX] < this_bbox[X0_IDX] < box[X1_IDX], box[X0_IDX] < this_bbox[X1_IDX] < box[X1_IDX],
this_bbox[X0_IDX] < box[X0_IDX] < this_bbox[X1_IDX], this_bbox[X0_IDX] < box[X1_IDX] < this_bbox[X1_IDX],
box[X0_IDX]==this_bbox[X0_IDX] and box[X1_IDX]==this_bbox[X1_IDX]])]
if len(top_bboxes)>0:
top_bboxes.sort(key=lambda x: x[Y1_EXT_IDX] if x[Y1_EXT_IDX] else x[Y1_IDX], reverse=True)
top_bboxes = top_bboxes[0]
else:
top_bboxes = None
return top_bboxes
def find_all_bottom_bbox_direct(this_bbox, all_bboxes) -> list:
"""
找到在this_bbox下侧且距离this_bbox距离最近的bbox.必须是直接遮挡的那种
"""
bottom_bboxes = [box for box in all_bboxes if box[Y0_IDX] >= this_bbox[Y1_IDX] and any([
this_bbox[X0_IDX] < box[X0_IDX] < this_bbox[X1_IDX], this_bbox[X0_IDX] < box[X1_IDX] < this_bbox[X1_IDX],
box[X0_IDX] < this_bbox[X0_IDX] < box[X1_IDX], box[X0_IDX] < this_bbox[X1_IDX] < box[X1_IDX],
box[X0_IDX]==this_bbox[X0_IDX] and box[X1_IDX]==this_bbox[X1_IDX]])]
if len(bottom_bboxes)>0:
bottom_bboxes.sort(key=lambda x: x[Y0_IDX])
bottom_bboxes = bottom_bboxes[0]
else:
bottom_bboxes = None
return bottom_bboxes
# ===================================================================================================================
def find_bottom_bbox_direct_from_right_edge(this_bbox, all_bboxes) -> list:
"""
找到在this_bbox下侧且距离this_bbox距离最近的bbox.必须是直接遮挡的那种
"""
bottom_bboxes = [box for box in all_bboxes if box[Y0_IDX] >= this_bbox[Y1_IDX] and any([
this_bbox[X0_IDX] < box[X0_IDX] < this_bbox[X1_IDX], this_bbox[X0_IDX] < box[X1_IDX] < this_bbox[X1_IDX],
box[X0_IDX] < this_bbox[X0_IDX] < box[X1_IDX], box[X0_IDX] < this_bbox[X1_IDX] < box[X1_IDX],
box[X0_IDX]==this_bbox[X0_IDX] and box[X1_IDX]==this_bbox[X1_IDX]])]
if len(bottom_bboxes)>0:
# y0最小， X1最大的那个,也就是box上边缘最靠近this_bbox的那个,并且还最靠右
bottom_bboxes.sort(key=lambda x: x[Y0_IDX])
bottom_bboxes = [box for box in bottom_bboxes if box[Y0_IDX]==bottom_bboxes[0][Y0_IDX]]
# 然后再y1相同的情况下，找到x1最大的那个
bottom_bboxes.sort(key=lambda x: x[X1_IDX], reverse=True)
bottom_bboxes = bottom_bboxes[0]
else:
bottom_bboxes = None
return bottom_bboxes
def find_bottom_bbox_direct_from_left_edge(this_bbox, all_bboxes) -> list:
"""
找到在this_bbox下侧且距离this_bbox距离最近的bbox.必须是直接遮挡的那种
"""
bottom_bboxes = [box for box in all_bboxes if box[Y0_IDX] >= this_bbox[Y1_IDX] and any([
this_bbox[X0_IDX] < box[X0_IDX] < this_bbox[X1_IDX], this_bbox[X0_IDX] < box[X1_IDX] < this_bbox[X1_IDX],
box[X0_IDX] < this_bbox[X0_IDX] < box[X1_IDX], box[X0_IDX] < this_bbox[X1_IDX] < box[X1_IDX],
box[X0_IDX]==this_bbox[X0_IDX] and box[X1_IDX]==this_bbox[X1_IDX]])]
if len(bottom_bboxes)>0:
# y0最小， X0最小的那个
bottom_bboxes.sort(key=lambda x: x[Y0_IDX])
bottom_bboxes = [box for box in bottom_bboxes if box[Y0_IDX]==bottom_bboxes[0][Y0_IDX]]
# 然后再y0相同的情况下，找到x0最小的那个
bottom_bboxes.sort(key=lambda x: x[X0_IDX])
bottom_bboxes = bottom_bboxes[0]
else:
bottom_bboxes = None
return bottom_bboxes
def find_top_bbox_direct_from_left_edge(this_bbox, all_bboxes) -> list:
"""
找到在this_bbox上侧且距离this_bbox距离最近的bbox.必须是直接遮挡的那种
"""
top_bboxes = [box for box in all_bboxes if box[Y1_IDX] <= this_bbox[Y0_IDX] and any([
box[X0_IDX] < this_bbox[X0_IDX] < box[X1_IDX], box[X0_IDX] < this_bbox[X1_IDX] < box[X1_IDX],
this_bbox[X0_IDX] < box[X0_IDX] < this_bbox[X1_IDX], this_bbox[X0_IDX] < box[X1_IDX] < this_bbox[X1_IDX],
box[X0_IDX]==this_bbox[X0_IDX] and box[X1_IDX]==this_bbox[X1_IDX]])]
if len(top_bboxes)>0:
# y1最大， X0最小的那个
top_bboxes.sort(key=lambda x: x[Y1_IDX], reverse=True)
top_bboxes = [box for box in top_bboxes if box[Y1_IDX]==top_bboxes[0][Y1_IDX]]
# 然后再y1相同的情况下，找到x0最小的那个
top_bboxes.sort(key=lambda x: x[X0_IDX])
top_bboxes = top_bboxes[0]
else:
top_bboxes = None
return top_bboxes
def find_top_bbox_direct_from_right_edge(this_bbox, all_bboxes) -> list:
"""
找到在this_bbox上侧且距离this_bbox距离最近的bbox.必须是直接遮挡的那种
"""
top_bboxes = [box for box in all_bboxes if box[Y1_IDX] <= this_bbox[Y0_IDX] and any([
box[X0_IDX] < this_bbox[X0_IDX] < box[X1_IDX], box[X0_IDX] < this_bbox[X1_IDX] < box[X1_IDX],
this_bbox[X0_IDX] < box[X0_IDX] < this_bbox[X1_IDX], this_bbox[X0_IDX] < box[X1_IDX] < this_bbox[X1_IDX],
box[X0_IDX]==this_bbox[X0_IDX] and box[X1_IDX]==this_bbox[X1_IDX]])]
if len(top_bboxes)>0:
# y1最大， X1最大的那个
top_bboxes.sort(key=lambda x: x[Y1_IDX], reverse=True)
top_bboxes = [box for box in top_bboxes if box[Y1_IDX]==top_bboxes[0][Y1_IDX]]
# 然后再y1相同的情况下，找到x1最大的那个
top_bboxes.sort(key=lambda x: x[X1_IDX], reverse=True)
top_bboxes = top_bboxes[0]
else:
top_bboxes = None
return top_bboxes
# ===================================================================================================================
def get_left_edge_bboxes(all_bboxes) -> list:
"""
返回最左边的bbox
"""
left_bboxes = [box for box in all_bboxes if find_all_left_bbox_direct(box, all_bboxes) is None]
return left_bboxes
def get_right_edge_bboxes(all_bboxes) -> list:
"""
返回最右边的bbox
"""
right_bboxes = [box for box in all_bboxes if find_all_right_bbox_direct(box, all_bboxes) is None]
return right_bboxes
def fix_vertical_bbox_pos(bboxes:list):
"""
检查这批bbox在垂直方向是否有轻微的重叠，如果重叠了，就把重叠的bbox往下移动一点
在x方向上必须一个包含或者被包含，或者完全重叠，不能只有部分重叠
"""
bboxes.sort(key=lambda x: x[Y0_IDX]) # 从上向下排列
for i in range(0, len(bboxes)):
for j in range(i+1, len(bboxes)):
if _is_bottom_full_overlap(bboxes[i][:4], bboxes[j][:4]):
# 如果两个bbox有部分重叠，那么就把下面的bbox往下移动一点
bboxes[j][Y0_IDX] = bboxes[i][Y1_IDX] + 2 # 2是个经验值
break
return bboxes
magic_pdf/layout/layout_sort.py deleted 100644 → 0
View file @ 4df1eb74
"""对pdf上的box进行layout识别，并对内部组成的box进行排序."""
from loguru import logger
from magic_pdf.layout.bbox_sort import (CONTENT_IDX, CONTENT_TYPE_IDX,
X0_EXT_IDX, X0_IDX, X1_EXT_IDX, X1_IDX,
Y0_EXT_IDX, Y0_IDX, Y1_EXT_IDX, Y1_IDX,
paper_bbox_sort)
from magic_pdf.layout.layout_det_utils import (
find_all_bottom_bbox_direct, find_all_left_bbox_direct,
find_all_right_bbox_direct, find_all_top_bbox_direct,
find_bottom_bbox_direct_from_left_edge,
find_bottom_bbox_direct_from_right_edge,
find_top_bbox_direct_from_left_edge, find_top_bbox_direct_from_right_edge,
get_left_edge_bboxes, get_right_edge_bboxes)
from magic_pdf.libs.boxbase import get_bbox_in_boundary
LAYOUT_V = 'V'
LAYOUT_H = 'H'
LAYOUT_UNPROC = 'U'
LAYOUT_BAD = 'B'
def _is_single_line_text(bbox):
"""检查bbox里面的文字是否只有一行."""
return True # TODO
box_type = bbox[CONTENT_TYPE_IDX]
if box_type != 'text':
return False
paras = bbox[CONTENT_IDX]['paras']
text_content = ''
for para_id, para in paras.items(): # 拼装内部的段落文本
is_title = para['is_title']
if is_title != 0:
text_content += f"## {para['text']}"
else:
text_content += para['text']
text_content += '\n\n'
return bbox[CONTENT_TYPE_IDX] == 'text' and len(text_content.split('\n\n')) <= 1
def _horizontal_split(bboxes: list, boundary: tuple, avg_font_size=20) -> list:
"""
对bboxes进行水平切割
方法是：找到左侧和右侧都没有被直接遮挡的box，然后进行扩展，之后进行切割
return:
返回几个大的Layout区域 [[x0, y0, x1, y1, "h|u|v"], ], h代表水平，u代表未探测的，v代表垂直布局
"""
sorted_layout_blocks = [] # 这是要最终返回的值
bound_x0, bound_y0, bound_x1, bound_y1 = boundary
all_bboxes = get_bbox_in_boundary(bboxes, boundary)
# all_bboxes = paper_bbox_sort(all_bboxes, abs(bound_x1-bound_x0), abs(bound_y1-bound_x0)) # 大致拍下序, 这个是基于直接遮挡的。
"""
首先在水平方向上扩展独占一行的bbox
"""
last_h_split_line_y1 = bound_y0 # 记录下上次的水平分割线
for i, bbox in enumerate(all_bboxes):
left_nearest_bbox = find_all_left_bbox_direct(bbox, all_bboxes) # 非扩展线
right_nearest_bbox = find_all_right_bbox_direct(bbox, all_bboxes)
if left_nearest_bbox is None and right_nearest_bbox is None: # 独占一行
"""
然而，如果只是孤立的一行文字，那么就还要满足以下几个条件才可以：
1. bbox和中心线相交。或者
2. 上方或者下方也存在同类水平的独占一行的bbox。 或者
3. TODO 加强条件：这个bbox上方和下方是同一列column，那么就不能算作独占一行
"""
# 先检查这个bbox里是否只包含一行文字
# is_single_line = _is_single_line_text(bbox)
"""
这里有个点需要注意，当页面内容不是居中的时候，第一次调用传递的是page的boundary，这个时候mid_x就不是中心线了.
所以这里计算出最紧致的boundary，然后再计算mid_x
"""
boundary_real_x0, boundary_real_x1 = min(
[bbox[X0_IDX] for bbox in all_bboxes]
), max([bbox[X1_IDX] for bbox in all_bboxes])
mid_x = (boundary_real_x0 + boundary_real_x1) / 2
# 检查这个box是否内容在中心线有交
# 必须跨过去2个字符的宽度
is_cross_boundary_mid_line = (
min(mid_x - bbox[X0_IDX], bbox[X1_IDX] - mid_x) > avg_font_size * 2
)
"""
检查条件2
"""
is_belong_to_col = False
"""
检查是否能被上方col吸收，方法是：
1. 上方非空且不是独占一行的，并且
2. 从上个水平分割的最大y=y1开始到当前bbox,最左侧的bbox的[min_x0, max_x1],能够覆盖当前box的[x0, x1]
"""
"""
以迭代的方式向上找，查找范围是[bound_x0, last_h_sp, bound_x1, bbox[Y0_IDX]]
"""
# 先确定上方的y0, y0
b_y0, b_y1 = last_h_split_line_y1, bbox[Y0_IDX]
# 然后从box开始逐个向上找到所有与box在x上有交集的box
box_to_check = [bound_x0, b_y0, bound_x1, b_y1]
bbox_in_bound_check = get_bbox_in_boundary(all_bboxes, box_to_check)
bboxes_on_top = []
virtual_box = bbox
while True:
b_on_top = find_all_top_bbox_direct(virtual_box, bbox_in_bound_check)
if b_on_top is not None:
bboxes_on_top.append(b_on_top)
virtual_box = [
min([virtual_box[X0_IDX], b_on_top[X0_IDX]]),
min(virtual_box[Y0_IDX], b_on_top[Y0_IDX]),
max([virtual_box[X1_IDX], b_on_top[X1_IDX]]),
b_y1,
]
else:
break
# 随后确定这些box的最小x0, 最大x1
if len(bboxes_on_top) > 0 and len(bboxes_on_top) != len(
bbox_in_bound_check
): # virtual_box可能会膨胀到占满整个区域，这实际上就不能属于一个col了。
min_x0, max_x1 = virtual_box[X0_IDX], virtual_box[X1_IDX]
# 然后采用一种比较粗糙的方法，看min_x0，max_x1是否与位于[bound_x0, last_h_sp, bound_x1, bbox[Y0_IDX]]之间的box有相交
if not any(
[
b[X0_IDX] <= min_x0 - 1 <= b[X1_IDX]
or b[X0_IDX] <= max_x1 + 1 <= b[X1_IDX]
for b in bbox_in_bound_check
]
):
# 其上，下都不能被扩展成行，暂时只检查一下上方 TODO
top_nearest_bbox = find_all_top_bbox_direct(bbox, bboxes)
bottom_nearest_bbox = find_all_bottom_bbox_direct(bbox, bboxes)
if not any(
[
top_nearest_bbox is not None
and (
find_all_left_bbox_direct(top_nearest_bbox, bboxes)
is None
and find_all_right_bbox_direct(top_nearest_bbox, bboxes)
is None
),
bottom_nearest_bbox is not None
and (
find_all_left_bbox_direct(bottom_nearest_bbox, bboxes)
is None
and find_all_right_bbox_direct(
bottom_nearest_bbox, bboxes
)
is None
),
top_nearest_bbox is None or bottom_nearest_bbox is None,
]
):
is_belong_to_col = True
# 检查是否能被下方col吸收 TODO
"""
这里为什么没有is_cross_boundary_mid_line的条件呢？
确实有些杂志左右两栏宽度不是对称的。
"""
if not is_belong_to_col or is_cross_boundary_mid_line:
bbox[X0_EXT_IDX] = bound_x0
bbox[Y0_EXT_IDX] = bbox[Y0_IDX]
bbox[X1_EXT_IDX] = bound_x1
bbox[Y1_EXT_IDX] = bbox[Y1_IDX]
last_h_split_line_y1 = bbox[Y1_IDX] # 更新这条线
else:
continue
"""
此时独占一行的被成功扩展到指定的边界上，这个时候利用边界条件合并连续的bbox，成为一个group
然后合并所有连续水平方向的bbox.
"""
all_bboxes.sort(key=lambda x: x[Y0_IDX])
h_bboxes = []
h_bbox_group = []
for bbox in all_bboxes:
if bbox[X0_EXT_IDX] == bound_x0 and bbox[X1_EXT_IDX] == bound_x1:
h_bbox_group.append(bbox)
else:
if len(h_bbox_group) > 0:
h_bboxes.append(h_bbox_group)
h_bbox_group = []
# 最后一个group
if len(h_bbox_group) > 0:
h_bboxes.append(h_bbox_group)
"""
现在h_bboxes里面是所有的group了，每个group都是一个list
对h_bboxes里的每个group进行计算放回到sorted_layouts里
"""
h_layouts = []
for gp in h_bboxes:
gp.sort(key=lambda x: x[Y0_IDX])
# 然后计算这个group的layout_bbox，也就是最小的x0,y0, 最大的x1,y1
x0, y0, x1, y1 = (
gp[0][X0_EXT_IDX],
gp[0][Y0_EXT_IDX],
gp[-1][X1_EXT_IDX],
gp[-1][Y1_EXT_IDX],
)
h_layouts.append([x0, y0, x1, y1, LAYOUT_H]) # 水平的布局
"""
接下来利用这些连续的水平bbox的layout_bbox的y0, y1，从水平上切分开其余的为几个部分
"""
h_split_lines = [bound_y0]
for gp in h_bboxes: # gp是一个list[bbox_list]
y0, y1 = gp[0][1], gp[-1][3]
h_split_lines.append(y0)
h_split_lines.append(y1)
h_split_lines.append(bound_y1)
unsplited_bboxes = []
for i in range(0, len(h_split_lines), 2):
start_y0, start_y1 = h_split_lines[i : i + 2]
# 然后找出[start_y0, start_y1]之间的其他bbox，这些组成一个未分割板块
bboxes_in_block = [
bbox
for bbox in all_bboxes
if bbox[Y0_IDX] >= start_y0 and bbox[Y1_IDX] <= start_y1
]
unsplited_bboxes.append(bboxes_in_block)
# 接着把未处理的加入到h_layouts里
for bboxes_in_block in unsplited_bboxes:
if len(bboxes_in_block) == 0:
continue
x0, y0, x1, y1 = (
bound_x0,
min([bbox[Y0_IDX] for bbox in bboxes_in_block]),
bound_x1,
max([bbox[Y1_IDX] for bbox in bboxes_in_block]),
)
h_layouts.append([x0, y0, x1, y1, LAYOUT_UNPROC])
h_layouts.sort(key=lambda x: x[1]) # 按照y0排序, 也就是从上到下的顺序
"""
转换成如下格式返回
"""
for layout in h_layouts:
sorted_layout_blocks.append(
{
'layout_bbox': layout[:4],
'layout_label': layout[4],
'sub_layout': [],
}
)
return sorted_layout_blocks
###############################################################################################
#
# 垂直方向的处理
#
#
###############################################################################################
def _vertical_align_split_v1(bboxes: list, boundary: tuple) -> list:
"""
计算垂直方向上的对齐， 并分割bboxes成layout。负责对一列多行的进行列维度分割。
如果不能完全分割，剩余部分作为layout_lable为u的layout返回
-----------------------
| | |
| | |
| | |
| | |
-------------------------
此函数会将：以上布局将会切分出来2列
"""
sorted_layout_blocks = [] # 这是要最终返回的值
new_boundary = [boundary[0], boundary[1], boundary[2], boundary[3]]
v_blocks = []
"""
先从左到右切分
"""
while True:
all_bboxes = get_bbox_in_boundary(bboxes, new_boundary)
left_edge_bboxes = get_left_edge_bboxes(all_bboxes)
if len(left_edge_bboxes) == 0:
break
right_split_line_x1 = max([bbox[X1_IDX] for bbox in left_edge_bboxes]) + 1
# 然后检查这条线能不与其他bbox的左边界相交或者重合
if any(
[bbox[X0_IDX] <= right_split_line_x1 <= bbox[X1_IDX] for bbox in all_bboxes]
):
# 垂直切分线与某些box发生相交，说明无法完全垂直方向切分。
break
else: # 说明成功分割出一列
# 找到左侧边界最靠左的bbox作为layout的x0
layout_x0 = min(
[bbox[X0_IDX] for bbox in left_edge_bboxes]
) # 这里主要是为了画出来有一定间距
v_blocks.append(
[
layout_x0,
new_boundary[1],
right_split_line_x1,
new_boundary[3],
LAYOUT_V,
]
)
new_boundary[0] = right_split_line_x1 # 更新边界
"""
再从右到左切， 此时如果还是无法完全切分，那么剩余部分作为layout_lable为u的layout返回
"""
unsplited_block = []
while True:
all_bboxes = get_bbox_in_boundary(bboxes, new_boundary)
right_edge_bboxes = get_right_edge_bboxes(all_bboxes)
if len(right_edge_bboxes) == 0:
break
left_split_line_x0 = min([bbox[X0_IDX] for bbox in right_edge_bboxes]) - 1
# 然后检查这条线能不与其他bbox的左边界相交或者重合
if any(
[bbox[X0_IDX] <= left_split_line_x0 <= bbox[X1_IDX] for bbox in all_bboxes]
):
# 这里是余下的
unsplited_block.append(
[
new_boundary[0],
new_boundary[1],
new_boundary[2],
new_boundary[3],
LAYOUT_UNPROC,
]
)
break
else:
# 找到右侧边界最靠右的bbox作为layout的x1
layout_x1 = max([bbox[X1_IDX] for bbox in right_edge_bboxes])
v_blocks.append(
[
left_split_line_x0,
new_boundary[1],
layout_x1,
new_boundary[3],
LAYOUT_V,
]
)
new_boundary[2] = left_split_line_x0 # 更新右边界
"""
最后拼装成layout格式返回
"""
for block in v_blocks:
sorted_layout_blocks.append(
{
'layout_bbox': block[:4],
'layout_label': block[4],
'sub_layout': [],
}
)
for block in unsplited_block:
sorted_layout_blocks.append(
{
'layout_bbox': block[:4],
'layout_label': block[4],
'sub_layout': [],
}
)
# 按照x0排序
sorted_layout_blocks.sort(key=lambda x: x['layout_bbox'][0])
return sorted_layout_blocks
def _vertical_align_split_v2(bboxes: list, boundary: tuple) -> list:
"""改进的
_vertical_align_split算法，原算法会因为第二列的box由于左侧没有遮挡被认为是左侧的一部分，导致整个layout多列被识别为一列。
利用从左上角的box开始向下看的方法，不断扩展w_x0, w_x1，直到不能继续向下扩展，或者到达边界下边界。"""
sorted_layout_blocks = [] # 这是要最终返回的值
new_boundary = [boundary[0], boundary[1], boundary[2], boundary[3]]
bad_boxes = [] # 被割中的box
v_blocks = []
while True:
all_bboxes = get_bbox_in_boundary(bboxes, new_boundary)
if len(all_bboxes) == 0:
break
left_top_box = min(
all_bboxes, key=lambda x: (x[X0_IDX], x[Y0_IDX])
) # 这里应该加强，检查一下必须是在第一列的 TODO
start_box = [
left_top_box[X0_IDX],
left_top_box[Y0_IDX],
left_top_box[X1_IDX],
left_top_box[Y1_IDX],
]
w_x0, w_x1 = left_top_box[X0_IDX], left_top_box[X1_IDX]
"""
然后沿着这个box线向下找最近的那个box, 然后扩展w_x0, w_x1
扩展之后，宽度会增加，随后用x=w_x1来检测在边界内是否有box与相交，如果相交，那么就说明不能再扩展了。
当不能扩展的时候就要看是否到达下边界：
1. 达到，那么更新左边界继续分下一个列
2. 没有达到，那么此时开始从右侧切分进入下面的循环里
"""
while left_top_box is not None: # 向下去找
virtual_box = [w_x0, left_top_box[Y0_IDX], w_x1, left_top_box[Y1_IDX]]
left_top_box = find_bottom_bbox_direct_from_left_edge(
virtual_box, all_bboxes
)
if left_top_box:
w_x0, w_x1 = min(virtual_box[X0_IDX], left_top_box[X0_IDX]), max(
[virtual_box[X1_IDX], left_top_box[X1_IDX]]
)
# 万一这个初始的box在column中间，那么还要向上看
start_box = [
w_x0,
start_box[Y0_IDX],
w_x1,
start_box[Y1_IDX],
] # 扩展一下宽度更鲁棒
left_top_box = find_top_bbox_direct_from_left_edge(start_box, all_bboxes)
while left_top_box is not None: # 向上去找
virtual_box = [w_x0, left_top_box[Y0_IDX], w_x1, left_top_box[Y1_IDX]]
left_top_box = find_top_bbox_direct_from_left_edge(virtual_box, all_bboxes)
if left_top_box:
w_x0, w_x1 = min(virtual_box[X0_IDX], left_top_box[X0_IDX]), max(
[virtual_box[X1_IDX], left_top_box[X1_IDX]]
)
# 检查相交
if any([bbox[X0_IDX] <= w_x1 + 1 <= bbox[X1_IDX] for bbox in all_bboxes]):
for b in all_bboxes:
if b[X0_IDX] <= w_x1 + 1 <= b[X1_IDX]:
bad_boxes.append([b[X0_IDX], b[Y0_IDX], b[X1_IDX], b[Y1_IDX]])
break
else: # 说明成功分割出一列
v_blocks.append([w_x0, new_boundary[1], w_x1, new_boundary[3], LAYOUT_V])
new_boundary[0] = w_x1 # 更新边界
"""
接着开始从右上角的box扫描
"""
w_x0, w_x1 = 0, 0
unsplited_block = []
while True:
all_bboxes = get_bbox_in_boundary(bboxes, new_boundary)
if len(all_bboxes) == 0:
break
# 先找到X1最大的
bbox_list_sorted = sorted(
all_bboxes, key=lambda bbox: bbox[X1_IDX], reverse=True
)
# Then, find the boxes with the smallest Y0 value
bigest_x1 = bbox_list_sorted[0][X1_IDX]
boxes_with_bigest_x1 = [
bbox for bbox in bbox_list_sorted if bbox[X1_IDX] == bigest_x1
] # 也就是最靠右的那些
right_top_box = min(
boxes_with_bigest_x1, key=lambda bbox: bbox[Y0_IDX]
) # y0最小的那个
start_box = [
right_top_box[X0_IDX],
right_top_box[Y0_IDX],
right_top_box[X1_IDX],
right_top_box[Y1_IDX],
]
w_x0, w_x1 = right_top_box[X0_IDX], right_top_box[X1_IDX]
while right_top_box is not None:
virtual_box = [w_x0, right_top_box[Y0_IDX], w_x1, right_top_box[Y1_IDX]]
right_top_box = find_bottom_bbox_direct_from_right_edge(
virtual_box, all_bboxes
)
if right_top_box:
w_x0, w_x1 = min([w_x0, right_top_box[X0_IDX]]), max(
[w_x1, right_top_box[X1_IDX]]
)
# 在向上扫描
start_box = [
w_x0,
start_box[Y0_IDX],
w_x1,
start_box[Y1_IDX],
] # 扩展一下宽度更鲁棒
right_top_box = find_top_bbox_direct_from_right_edge(start_box, all_bboxes)
while right_top_box is not None:
virtual_box = [w_x0, right_top_box[Y0_IDX], w_x1, right_top_box[Y1_IDX]]
right_top_box = find_top_bbox_direct_from_right_edge(
virtual_box, all_bboxes
)
if right_top_box:
w_x0, w_x1 = min([w_x0, right_top_box[X0_IDX]]), max(
[w_x1, right_top_box[X1_IDX]]
)
# 检查是否与其他box相交， 垂直切分线与某些box发生相交，说明无法完全垂直方向切分。
if any([bbox[X0_IDX] <= w_x0 - 1 <= bbox[X1_IDX] for bbox in all_bboxes]):
unsplited_block.append(
[
new_boundary[0],
new_boundary[1],
new_boundary[2],
new_boundary[3],
LAYOUT_UNPROC,
]
)
for b in all_bboxes:
if b[X0_IDX] <= w_x0 - 1 <= b[X1_IDX]:
bad_boxes.append([b[X0_IDX], b[Y0_IDX], b[X1_IDX], b[Y1_IDX]])
break
else: # 说明成功分割出一列
v_blocks.append([w_x0, new_boundary[1], w_x1, new_boundary[3], LAYOUT_V])
new_boundary[2] = w_x0
"""转换数据结构"""
for block in v_blocks:
sorted_layout_blocks.append(
{
'layout_bbox': block[:4],
'layout_label': block[4],
'sub_layout': [],
}
)
for block in unsplited_block:
sorted_layout_blocks.append(
{
'layout_bbox': block[:4],
'layout_label': block[4],
'sub_layout': [],
'bad_boxes': bad_boxes, # 记录下来，这个box是被割中的
}
)
# 按照x0排序
sorted_layout_blocks.sort(key=lambda x: x['layout_bbox'][0])
return sorted_layout_blocks
def _try_horizontal_mult_column_split(bboxes: list, boundary: tuple) -> list:
"""
尝试水平切分，如果切分不动，那就当一个BAD_LAYOUT返回
------------------
| | |
------------------
| | | | <- 这里是此函数要切分的场景
------------------
| | |
| | |
"""
pass
def _vertical_split(bboxes: list, boundary: tuple) -> list:
"""
从垂直方向进行切割，分block
这个版本里，如果垂直切分不动，那就当一个BAD_LAYOUT返回
--------------------------
| | |
| | |
| |
这种列是此函数要切分的 -> | |
| |
| | |
| | |
-------------------------
"""
sorted_layout_blocks = [] # 这是要最终返回的值
bound_x0, bound_y0, bound_x1, bound_y1 = boundary
all_bboxes = get_bbox_in_boundary(bboxes, boundary)
"""
all_bboxes = fix_vertical_bbox_pos(all_bboxes) # 垂直方向解覆盖
all_bboxes = fix_hor_bbox_pos(all_bboxes) # 水平解覆盖
这两行代码目前先不执行，因为公式检测，表格检测还不是很成熟，导致非常多的textblock参与了运算，时间消耗太大。
这两行代码的作用是：
如果遇到互相重叠的bbox, 那么会把面积较小的box进行压缩，从而避免重叠。对布局切分来说带来正反馈。
"""
# all_bboxes = paper_bbox_sort(all_bboxes, abs(bound_x1-bound_x0), abs(bound_y1-bound_x0)) # 大致拍下序, 这个是基于直接遮挡的。
"""
首先在垂直方向上扩展独占一行的bbox
"""
for bbox in all_bboxes:
top_nearest_bbox = find_all_top_bbox_direct(bbox, all_bboxes) # 非扩展线
bottom_nearest_bbox = find_all_bottom_bbox_direct(bbox, all_bboxes)
if (
top_nearest_bbox is None
and bottom_nearest_bbox is None
and not any(
[
b[X0_IDX] < bbox[X1_IDX] < b[X1_IDX]
or b[X0_IDX] < bbox[X0_IDX] < b[X1_IDX]
for b in all_bboxes
]
)
): # 独占一列, 且不和其他重叠
bbox[X0_EXT_IDX] = bbox[X0_IDX]
bbox[Y0_EXT_IDX] = bound_y0
bbox[X1_EXT_IDX] = bbox[X1_IDX]
bbox[Y1_EXT_IDX] = bound_y1
"""
此时独占一列的被成功扩展到指定的边界上，这个时候利用边界条件合并连续的bbox，成为一个group
然后合并所有连续垂直方向的bbox.
"""
all_bboxes.sort(key=lambda x: x[X0_IDX])
# fix: 这里水平方向的列不要合并成一个行，因为需要保证返回给下游的最小block，总是可以无脑从上到下阅读文字。
v_bboxes = []
for box in all_bboxes:
if box[Y0_EXT_IDX] == bound_y0 and box[Y1_EXT_IDX] == bound_y1:
v_bboxes.append(box)
"""
现在v_bboxes里面是所有的group了，每个group都是一个list
对v_bboxes里的每个group进行计算放回到sorted_layouts里
"""
v_layouts = []
for vbox in v_bboxes:
# gp.sort(key=lambda x: x[X0_IDX])
# 然后计算这个group的layout_bbox，也就是最小的x0,y0, 最大的x1,y1
x0, y0, x1, y1 = (
vbox[X0_EXT_IDX],
vbox[Y0_EXT_IDX],
vbox[X1_EXT_IDX],
vbox[Y1_EXT_IDX],
)
v_layouts.append([x0, y0, x1, y1, LAYOUT_V]) # 垂直的布局
"""
接下来利用这些连续的垂直bbox的layout_bbox的x0, x1，从垂直上切分开其余的为几个部分
"""
v_split_lines = [bound_x0]
for gp in v_bboxes:
x0, x1 = gp[X0_IDX], gp[X1_IDX]
v_split_lines.append(x0)
v_split_lines.append(x1)
v_split_lines.append(bound_x1)
unsplited_bboxes = []
for i in range(0, len(v_split_lines), 2):
start_x0, start_x1 = v_split_lines[i : i + 2]
# 然后找出[start_x0, start_x1]之间的其他bbox，这些组成一个未分割板块
bboxes_in_block = [
bbox
for bbox in all_bboxes
if bbox[X0_IDX] >= start_x0 and bbox[X1_IDX] <= start_x1
]
unsplited_bboxes.append(bboxes_in_block)
# 接着把未处理的加入到v_layouts里
for bboxes_in_block in unsplited_bboxes:
if len(bboxes_in_block) == 0:
continue
x0, y0, x1, y1 = (
min([bbox[X0_IDX] for bbox in bboxes_in_block]),
bound_y0,
max([bbox[X1_IDX] for bbox in bboxes_in_block]),
bound_y1,
)
v_layouts.append(
[x0, y0, x1, y1, LAYOUT_UNPROC]
) # 说明这篇区域未能够分析出可靠的版面
v_layouts.sort(key=lambda x: x[0]) # 按照x0排序, 也就是从左到右的顺序
for layout in v_layouts:
sorted_layout_blocks.append(
{
'layout_bbox': layout[:4],
'layout_label': layout[4],
'sub_layout': [],
}
)
"""
至此，垂直方向切成了2种类型，其一是独占一列的，其二是未处理的。
下面对这些未处理的进行垂直方向切分，这个切分要切出来类似“吕”这种类型的垂直方向的布局
"""
for i, layout in enumerate(sorted_layout_blocks):
if layout['layout_label'] == LAYOUT_UNPROC:
x0, y0, x1, y1 = layout['layout_bbox']
v_split_layouts = _vertical_align_split_v2(bboxes, [x0, y0, x1, y1])
sorted_layout_blocks[i] = {
'layout_bbox': [x0, y0, x1, y1],
'layout_label': LAYOUT_H,
'sub_layout': v_split_layouts,
}
layout['layout_label'] = LAYOUT_H # 被垂线切分成了水平布局
return sorted_layout_blocks
def split_layout(bboxes: list, boundary: tuple, page_num: int) -> list:
"""
把bboxes切割成layout
return:
[
{
"layout_bbox": [x0,y0,x1,y1],
"layout_label":"u|v|h|b", 未处理|垂直|水平|BAD_LAYOUT
"sub_layout":[] #每个元素都是[
x0,y0,
x1,y1,
block_content,
idx_x,idx_y,
content_type,
ext_x0,ext_y0,
ext_x1,ext_y1
], 并且顺序就是阅读顺序
}
]
example:
[
{
"layout_bbox": [0, 0, 100, 100],
"layout_label":"u|v|h|b",
"sub_layout":[
]
},
{
"layout_bbox": [0, 0, 100, 100],
"layout_label":"u|v|h|b",
"sub_layout":[
{
"layout_bbox": [0, 0, 100, 100],
"layout_label":"u|v|h|b",
"content_bboxes":[
[],
[],
[]
]
},
{
"layout_bbox": [0, 0, 100, 100],
"layout_label":"u|v|h|b",
"sub_layout":[
]
}
}
]
"""
sorted_layouts = [] # 最终返回的结果
boundary_x0, boundary_y0, boundary_x1, boundary_y1 = boundary
if len(bboxes) <= 1:
return [
{
'layout_bbox': [boundary_x0, boundary_y0, boundary_x1, boundary_y1],
'layout_label': LAYOUT_V,
'sub_layout': [],
}
]
"""
接下来按照先水平后垂直的顺序进行切分
"""
bboxes = paper_bbox_sort(
bboxes, boundary_x1 - boundary_x0, boundary_y1 - boundary_y0
)
sorted_layouts = _horizontal_split(bboxes, boundary) # 通过水平分割出来的layout
for i, layout in enumerate(sorted_layouts):
x0, y0, x1, y1 = layout['layout_bbox']
layout_type = layout['layout_label']
if layout_type == LAYOUT_UNPROC: # 说明是非独占单行的，这些需要垂直切分
v_split_layouts = _vertical_split(bboxes, [x0, y0, x1, y1])
"""
最后这里有个逻辑问题：如果这个函数只分离出来了一个column layout，那么这个layout分割肯定超出了算法能力范围。因为我们假定的是传进来的
box已经把行全部剥离了，所以这里必须十多个列才可以。如果只剥离出来一个layout，并且是多个box，那么就说明这个layout是无法分割的，标记为LAYOUT_UNPROC
"""
layout_label = LAYOUT_V
if len(v_split_layouts) == 1:
if len(v_split_layouts[0]['sub_layout']) == 0:
layout_label = LAYOUT_UNPROC
# logger.warning(f"WARNING: pageno={page_num}, 无法分割的layout: ", v_split_layouts)
"""
组合起来最终的layout
"""
sorted_layouts[i] = {
'layout_bbox': [x0, y0, x1, y1],
'layout_label': layout_label,
'sub_layout': v_split_layouts,
}
layout['layout_label'] = LAYOUT_H
"""
水平和垂直方向都切分完毕了。此时还有一些未处理的，这些未处理的可能是因为水平和垂直方向都无法切分。
这些最后调用_try_horizontal_mult_block_split做一次水平多个block的联合切分，如果也不能切分最终就当做BAD_LAYOUT返回
"""
# TODO
return sorted_layouts
def get_bboxes_layout(all_boxes: list, boundary: tuple, page_id: int):
"""
对利用layout排序之后的box，进行排序
return:
[
{
"layout_bbox": [x0, y0, x1, y1],
"layout_label":"u|v|h|b", 未处理|垂直|水平|BAD_LAYOUT
}，
]
"""
def _preorder_traversal(layout):
"""对sorted_layouts的叶子节点，也就是len(sub_layout)==0的节点进行排序。排序按照前序遍历的顺序，也就是从上到
下，从左到右的顺序."""
sorted_layout_blocks = []
for layout in layout:
sub_layout = layout['sub_layout']
if len(sub_layout) == 0:
sorted_layout_blocks.append(layout)
else:
s = _preorder_traversal(sub_layout)
sorted_layout_blocks.extend(s)
return sorted_layout_blocks
# -------------------------------------------------------------------------------------------------------------------------
sorted_layouts = split_layout(
all_boxes, boundary, page_id
) # 先切分成layout，得到一个Tree
total_sorted_layout_blocks = _preorder_traversal(sorted_layouts)
return total_sorted_layout_blocks, sorted_layouts
def get_columns_cnt_of_layout(layout_tree):
"""获取一个layout的宽度."""
max_width_list = [0] # 初始化一个元素，防止max,min函数报错
for items in layout_tree: # 针对每一层（横切）计算列数，横着的算一列
layout_type = items['layout_label']
sub_layouts = items['sub_layout']
if len(sub_layouts) == 0:
max_width_list.append(1)
else:
if layout_type == LAYOUT_H:
max_width_list.append(1)
else:
width = 0
for sub_layout in sub_layouts:
if len(sub_layout['sub_layout']) == 0:
width += 1
else:
for lay in sub_layout['sub_layout']:
width += get_columns_cnt_of_layout([lay])
max_width_list.append(width)
return max(max_width_list)
def sort_with_layout(bboxes: list, page_width, page_height) -> (list, list):
"""输入是一个bbox的list.
获取到输入之后，先进行layout切分，然后对这些bbox进行排序。返回排序后的bboxes
"""
new_bboxes = []
for box in bboxes:
# new_bboxes.append([box[0], box[1], box[2], box[3], None, None, None, 'text', None, None, None, None])
new_bboxes.append(
[
box[0],
box[1],
box[2],
box[3],
None,
None,
None,
'text',
None,
None,
None,
None,
box[4],
]
)
layout_bboxes, _ = get_bboxes_layout(
new_bboxes, tuple([0, 0, page_width, page_height]), 0
)
if any([lay['layout_label'] == LAYOUT_UNPROC for lay in layout_bboxes]):
logger.warning('drop this pdf, reason: 复杂版面')
return None, None
sorted_bboxes = []
# 利用layout bbox每次框定一些box，然后排序
for layout in layout_bboxes:
lbox = layout['layout_bbox']
bbox_in_layout = get_bbox_in_boundary(new_bboxes, lbox)
sorted_bbox = paper_bbox_sort(
bbox_in_layout, lbox[2] - lbox[0], lbox[3] - lbox[1]
)
sorted_bboxes.extend(sorted_bbox)
return sorted_bboxes, layout_bboxes
def sort_text_block(text_block, layout_bboxes):
"""对一页的text_block进行排序."""
sorted_text_bbox = []
all_text_bbox = []
# 做一个box=>text的映射
box_to_text = {}
for blk in text_block:
box = blk['bbox']
box_to_text[(box[0], box[1], box[2], box[3])] = blk
all_text_bbox.append(box)
# text_blocks_to_sort = []
# for box in box_to_text.keys():
# text_blocks_to_sort.append([box[0], box[1], box[2], box[3], None, None, None, 'text', None, None, None, None])
# 按照layout_bboxes的顺序，对text_block进行排序
for layout in layout_bboxes:
layout_box = layout['layout_bbox']
text_bbox_in_layout = get_bbox_in_boundary(
all_text_bbox,
[
layout_box[0] - 1,
layout_box[1] - 1,
layout_box[2] + 1,
layout_box[3] + 1,
],
)
# sorted_bbox = paper_bbox_sort(text_bbox_in_layout, layout_box[2]-layout_box[0], layout_box[3]-layout_box[1])
text_bbox_in_layout.sort(
key=lambda x: x[1]
) # 一个layout内部的box，按照y0自上而下排序
# sorted_bbox = [[b] for b in text_blocks_to_sort]
for sb in text_bbox_in_layout:
sorted_text_bbox.append(box_to_text[(sb[0], sb[1], sb[2], sb[3])])
return sorted_text_bbox
magic_pdf/layout/layout_spiler_recog.py deleted 100644 → 0
View file @ 4df1eb74
"""
找到能分割布局的水平的横线、色块
"""
import os
from magic_pdf.libs.commons import fitz
from magic_pdf.libs.boxbase import _is_in_or_part_overlap
def __rect_filter_by_width(rect, page_w, page_h):
mid_x = page_w/2
if rect[0]< mid_x < rect[2]:
return True
return False
def __rect_filter_by_pos(rect, image_bboxes, table_bboxes):
"""
不能出现在table和image的位置
"""
for box in image_bboxes:
if _is_in_or_part_overlap(rect, box):
return False
for box in table_bboxes:
if _is_in_or_part_overlap(rect, box):
return False
return True
def __debug_show_page(page, bboxes1: list,bboxes2: list,bboxes3: list,):
save_path = "./tmp/debug.pdf"
if os.path.exists(save_path):
# 删除已经存在的文件
os.remove(save_path)
# 创建一个新的空白 PDF 文件
doc = fitz.open('')
width = page.rect.width
height = page.rect.height
new_page = doc.new_page(width=width, height=height)
shape = new_page.new_shape()
for bbox in bboxes1:
# 原始box画上去
rect = fitz.Rect(*bbox[0:4])
shape = new_page.new_shape()
shape.draw_rect(rect)
shape.finish(color=fitz.pdfcolor['red'], fill=fitz.pdfcolor['blue'], fill_opacity=0.2)
shape.finish()
shape.commit()
for bbox in bboxes2:
# 原始box画上去
rect = fitz.Rect(*bbox[0:4])
shape = new_page.new_shape()
shape.draw_rect(rect)
shape.finish(color=None, fill=fitz.pdfcolor['yellow'], fill_opacity=0.2)
shape.finish()
shape.commit()
for bbox in bboxes3:
# 原始box画上去
rect = fitz.Rect(*bbox[0:4])
shape = new_page.new_shape()
shape.draw_rect(rect)
shape.finish(color=fitz.pdfcolor['red'], fill=None)
shape.finish()
shape.commit()
parent_dir = os.path.dirname(save_path)
if not os.path.exists(parent_dir):
os.makedirs(parent_dir)
doc.save(save_path)
doc.close()
def get_spilter_of_page(page, image_bboxes, table_bboxes):
"""
获取到色块和横线
"""
cdrawings = page.get_cdrawings()
spilter_bbox = []
for block in cdrawings:
if 'fill' in block:
fill = block['fill']
if 'fill' in block and block['fill'] and block['fill']!=(1.0,1.0,1.0):
rect = block['rect']
if __rect_filter_by_width(rect, page.rect.width, page.rect.height) and __rect_filter_by_pos(rect, image_bboxes, table_bboxes):
spilter_bbox.append(list(rect))
"""过滤、修正一下这些box。因为有时候会有一些矩形，高度为0或者为负数，造成layout计算无限循环。如果是负高度或者0高度，统一修正为高度为1"""
for box in spilter_bbox:
if box[3]-box[1] <= 0:
box[3] = box[1] + 1
#__debug_show_page(page, spilter_bbox, [], [])
return spilter_bbox
magic_pdf/layout/mcol_sort.py deleted 100644 → 0
View file @ 4df1eb74
"""
This is an advanced PyMuPDF utility for detecting multi-column pages.
It can be used in a shell script, or its main function can be imported and
invoked as descript below.
Features
---------
- Identify text belonging to (a variable number of) columns on the page.
- Text with different background color is handled separately, allowing for
easier treatment of side remarks, comment boxes, etc.
- Uses text block detection capability to identify text blocks and
uses the block bboxes as primary structuring principle.
- Supports ignoring footers via a footer margin parameter.
- Returns re-created text boundary boxes (integer coordinates), sorted ascending
by the top, then by the left coordinates.
Restrictions
-------------
- Only supporting horizontal, left-to-right text
- Returns a list of text boundary boxes - not the text itself. The caller is
expected to extract text from within the returned boxes.
- Text written above images is ignored altogether (option).
- This utility works as expected in most cases. The following situation cannot
be handled correctly:
* overlapping (non-disjoint) text blocks
* image captions are not recognized and are handled like normal text
Usage
------
- As a CLI shell command use
python multi_column.py input.pdf footer_margin
Where footer margin is the height of the bottom stripe to ignore on each page.
This code is intended to be modified according to your need.
- Use in a Python script as follows:
----------------------------------------------------------------------------------
from multi_column import column_boxes
# for each page execute
bboxes = column_boxes(page, footer_margin=50, no_image_text=True)
# bboxes is a list of fitz.IRect objects, that are sort ascending by their y0,
# then x0 coordinates. Their text content can be extracted by all PyMuPDF
# get_text() variants, like for instance the following:
for rect in bboxes:
print(page.get_text(clip=rect, sort=True))
----------------------------------------------------------------------------------
"""
import sys
from magic_pdf.libs.commons import fitz
def column_boxes(page, footer_margin=50, header_margin=50, no_image_text=True):
"""Determine bboxes which wrap a column."""
paths = page.get_drawings()
bboxes = []
# path rectangles
path_rects = []
# image bboxes
img_bboxes = []
# bboxes of non-horizontal text
# avoid when expanding horizontal text boxes
vert_bboxes = []
# compute relevant page area
clip = +page.rect
clip.y1 -= footer_margin # Remove footer area
clip.y0 += header_margin # Remove header area
def can_extend(temp, bb, bboxlist):
"""Determines whether rectangle 'temp' can be extended by 'bb'
without intersecting any of the rectangles contained in 'bboxlist'.
Items of bboxlist may be None if they have been removed.
Returns:
True if 'temp' has no intersections with items of 'bboxlist'.
"""
for b in bboxlist:
if not intersects_bboxes(temp, vert_bboxes) and (
b == None or b == bb or (temp & b).is_empty
):
continue
return False
return True
def in_bbox(bb, bboxes):
"""Return 1-based number if a bbox contains bb, else return 0."""
for i, bbox in enumerate(bboxes):
if bb in bbox:
return i + 1
return 0
def intersects_bboxes(bb, bboxes):
"""Return True if a bbox intersects bb, else return False."""
for bbox in bboxes:
if not (bb & bbox).is_empty:
return True
return False
def extend_right(bboxes, width, path_bboxes, vert_bboxes, img_bboxes):
"""Extend a bbox to the right page border.
Whenever there is no text to the right of a bbox, enlarge it up
to the right page border.
Args:
bboxes: (list[IRect]) bboxes to check
width: (int) page width
path_bboxes: (list[IRect]) bboxes with a background color
vert_bboxes: (list[IRect]) bboxes with vertical text
img_bboxes: (list[IRect]) bboxes of images
Returns:
Potentially modified bboxes.
"""
for i, bb in enumerate(bboxes):
# do not extend text with background color
if in_bbox(bb, path_bboxes):
continue
# do not extend text in images
if in_bbox(bb, img_bboxes):
continue
# temp extends bb to the right page border
temp = +bb
temp.x1 = width
# do not cut through colored background or images
if intersects_bboxes(temp, path_bboxes + vert_bboxes + img_bboxes):
continue
# also, do not intersect other text bboxes
check = can_extend(temp, bb, bboxes)
if check:
bboxes[i] = temp # replace with enlarged bbox
return [b for b in bboxes if b != None]
def clean_nblocks(nblocks):
"""Do some elementary cleaning."""
# 1. remove any duplicate blocks.
blen = len(nblocks)
if blen < 2:
return nblocks
start = blen - 1
for i in range(start, -1, -1):
bb1 = nblocks[i]
bb0 = nblocks[i - 1]
if bb0 == bb1:
del nblocks[i]
# 2. repair sequence in special cases:
# consecutive bboxes with almost same bottom value are sorted ascending
# by x-coordinate.
y1 = nblocks[0].y1 # first bottom coordinate
i0 = 0 # its index
i1 = -1 # index of last bbox with same bottom
# Iterate over bboxes, identifying segments with approx. same bottom value.
# Replace every segment by its sorted version.
for i in range(1, len(nblocks)):
b1 = nblocks[i]
if abs(b1.y1 - y1) > 10: # different bottom
if i1 > i0: # segment length > 1? Sort it!
nblocks[i0 : i1 + 1] = sorted(
nblocks[i0 : i1 + 1], key=lambda b: b.x0
)
y1 = b1.y1 # store new bottom value
i0 = i # store its start index
i1 = i # store current index
if i1 > i0: # segment waiting to be sorted
nblocks[i0 : i1 + 1] = sorted(nblocks[i0 : i1 + 1], key=lambda b: b.x0)
return nblocks
# extract vector graphics
for p in paths:
path_rects.append(p["rect"].irect)
path_bboxes = path_rects
# sort path bboxes by ascending top, then left coordinates
path_bboxes.sort(key=lambda b: (b.y0, b.x0))
# bboxes of images on page, no need to sort them
for item in page.get_images():
img_bboxes.extend(page.get_image_rects(item[0]))
# blocks of text on page
blocks = page.get_text(
"dict",
flags=fitz.TEXTFLAGS_TEXT,
clip=clip,
)["blocks"]
# Make block rectangles, ignoring non-horizontal text
for b in blocks:
bbox = fitz.IRect(b["bbox"]) # bbox of the block
# ignore text written upon images
if no_image_text and in_bbox(bbox, img_bboxes):
continue
# confirm first line to be horizontal
line0 = b["lines"][0] # get first line
if line0["dir"] != (1, 0): # only accept horizontal text
vert_bboxes.append(bbox)
continue
srect = fitz.EMPTY_IRECT()
for line in b["lines"]:
lbbox = fitz.IRect(line["bbox"])
text = "".join([s["text"].strip() for s in line["spans"]])
if len(text) > 1:
srect |= lbbox
bbox = +srect
if not bbox.is_empty:
bboxes.append(bbox)
# Sort text bboxes by ascending background, top, then left coordinates
bboxes.sort(key=lambda k: (in_bbox(k, path_bboxes), k.y0, k.x0))
# Extend bboxes to the right where possible
bboxes = extend_right(
bboxes, int(page.rect.width), path_bboxes, vert_bboxes, img_bboxes
)
# immediately return of no text found
if bboxes == []:
return []
# --------------------------------------------------------------------
# Join bboxes to establish some column structure
# --------------------------------------------------------------------
# the final block bboxes on page
nblocks = [bboxes[0]] # pre-fill with first bbox
bboxes = bboxes[1:] # remaining old bboxes
for i, bb in enumerate(bboxes): # iterate old bboxes
check = False # indicates unwanted joins
# check if bb can extend one of the new blocks
for j in range(len(nblocks)):
nbb = nblocks[j] # a new block
# never join across columns
if bb == None or nbb.x1 < bb.x0 or bb.x1 < nbb.x0:
continue
# never join across different background colors
if in_bbox(nbb, path_bboxes) != in_bbox(bb, path_bboxes):
continue
temp = bb | nbb # temporary extension of new block
check = can_extend(temp, nbb, nblocks)
if check == True:
break
if not check: # bb cannot be used to extend any of the new bboxes
nblocks.append(bb) # so add it to the list
j = len(nblocks) - 1 # index of it
temp = nblocks[j] # new bbox added
# check if some remaining bbox is contained in temp
check = can_extend(temp, bb, bboxes)
if check == False:
nblocks.append(bb)
else:
nblocks[j] = temp
bboxes[i] = None
# do some elementary cleaning
nblocks = clean_nblocks(nblocks)
# return identified text bboxes
return nblocks
if __name__ == "__main__":
"""Only for debugging purposes, currently.
Draw red borders around the returned text bboxes and insert
the bbox number.
Then save the file under the name "input-blocks.pdf".
"""
# get the file name
filename = sys.argv[1]
# check if footer margin is given
if len(sys.argv) > 2:
footer_margin = int(sys.argv[2])
else: # use default vaue
footer_margin = 50
# check if header margin is given
if len(sys.argv) > 3:
header_margin = int(sys.argv[3])
else: # use default vaue
header_margin = 50
# open document
doc = fitz.open(filename)
# iterate over the pages
for page in doc:
# remove any geometry issues
page.wrap_contents()
# get the text bboxes
bboxes = column_boxes(page, footer_margin=footer_margin, header_margin=header_margin)
# prepare a canvas to draw rectangles and text
shape = page.new_shape()
# iterate over the bboxes
for i, rect in enumerate(bboxes):
shape.draw_rect(rect) # draw a border
# write sequence number
shape.insert_text(rect.tl + (5, 15), str(i), color=fitz.pdfcolor["red"])
# finish drawing / text with color red
shape.finish(color=fitz.pdfcolor["red"])
shape.commit() # store to the page
# save document with text bboxes
doc.ez_save(filename.replace(".pdf", "-blocks.pdf"))
\ No newline at end of file
magic_pdf/libs/calc_span_stats.py deleted 100644 → 0
View file @ 4df1eb74
import os
import csv
import json
import pandas as pd
from pandas import DataFrame as df
from matplotlib import pyplot as plt
from termcolor import cprint
"""
Execute this script in the following way:
1. Make sure there are pdf_dic.json files under the directory code-clean/tmp/unittest/md/, such as the following:
code-clean/tmp/unittest/md/scihub/scihub_00500000/libgen.scimag00527000-00527999.zip_10.1002/app.25178/pdf_dic.json
2. Under the directory code-clean, execute the following command:
$ python -m libs.calc_span_stats
"""
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 safe_get(dict_obj, key, default):
val = dict_obj.get(key)
if val is None:
return default
else:
return val
class SpanStatsCalc:
"""Calculate statistics of span."""
def draw_charts(self, span_stats: pd.DataFrame, fig_num: int, save_path: str):
"""Draw multiple figures in one figure."""
# make a canvas
fig = plt.figure(fig_num, figsize=(20, 20))
pass
def calc_stats_per_dict(self, pdf_dict) -> pd.DataFrame:
"""Calculate statistics per pdf_dict."""
span_stats = pd.DataFrame()
span_stats = []
span_id = 0
for page_id, blocks in pdf_dict.items():
if page_id.startswith("page_"):
if "para_blocks" in blocks.keys():
for para_block in blocks["para_blocks"]:
for line in para_block["lines"]:
for span in line["spans"]:
span_text = safe_get(span, "text", "")
span_font_name = safe_get(span, "font", "")
span_font_size = safe_get(span, "size", 0)
span_font_color = safe_get(span, "color", "")
span_font_flags = safe_get(span, "flags", 0)
span_font_flags_decoded = safe_get(span, "decomposed_flags", {})
span_is_super_script = safe_get(span_font_flags_decoded, "is_superscript", False)
span_is_italic = safe_get(span_font_flags_decoded, "is_italic", False)
span_is_serifed = safe_get(span_font_flags_decoded, "is_serifed", False)
span_is_sans_serifed = safe_get(span_font_flags_decoded, "is_sans_serifed", False)
span_is_monospaced = safe_get(span_font_flags_decoded, "is_monospaced", False)
span_is_proportional = safe_get(span_font_flags_decoded, "is_proportional", False)
span_is_bold = safe_get(span_font_flags_decoded, "is_bold", False)
span_stats.append(
{
"span_id": span_id, # id of span
"page_id": page_id, # page number of pdf
"span_text": span_text, # text of span
"span_font_name": span_font_name, # font name of span
"span_font_size": span_font_size, # font size of span
"span_font_color": span_font_color, # font color of span
"span_font_flags": span_font_flags, # font flags of span
"span_is_superscript": int(
span_is_super_script
), # indicate whether the span is super script or not
"span_is_italic": int(span_is_italic), # indicate whether the span is italic or not
"span_is_serifed": int(span_is_serifed), # indicate whether the span is serifed or not
"span_is_sans_serifed": int(
span_is_sans_serifed
), # indicate whether the span is sans serifed or not
"span_is_monospaced": int(
span_is_monospaced
), # indicate whether the span is monospaced or not
"span_is_proportional": int(
span_is_proportional
), # indicate whether the span is proportional or not
"span_is_bold": int(span_is_bold), # indicate whether the span is bold or not
}
)
span_id += 1
span_stats = pd.DataFrame(span_stats)
# print(span_stats)
return span_stats
def __find_pdf_dic_files(
jf_name="pdf_dic.json",
base_code_name="code-clean",
tgt_base_dir_name="tmp",
unittest_dir_name="unittest",
md_dir_name="md",
book_names=[
"scihub",
], # other possible values: "zlib", "arxiv" and so on
):
pdf_dict_files = []
curr_dir = os.path.dirname(__file__)
for i in range(len(curr_dir)):
if curr_dir[i : i + len(base_code_name)] == base_code_name:
base_code_dir_name = curr_dir[: i + len(base_code_name)]
for book_name in book_names:
search_dir_relative_name = os.path.join(tgt_base_dir_name, unittest_dir_name, md_dir_name, book_name)
if os.path.exists(base_code_dir_name):
search_dir_name = os.path.join(base_code_dir_name, search_dir_relative_name)
for root, dirs, files in os.walk(search_dir_name):
for file in files:
if file == jf_name:
pdf_dict_files.append(os.path.join(root, file))
break
return pdf_dict_files
def combine_span_texts(group_df, span_stats):
combined_span_texts = []
for _, row in group_df.iterrows():
curr_span_id = row.name
curr_span_text = row["span_text"]
pre_span_id = curr_span_id - 1
pre_span_text = span_stats.at[pre_span_id, "span_text"] if pre_span_id in span_stats.index else ""
next_span_id = curr_span_id + 1
next_span_text = span_stats.at[next_span_id, "span_text"] if next_span_id in span_stats.index else ""
# pointer_sign is a right arrow if the span is superscript, otherwise it is a down arrow
pointer_sign = "→ → → "
combined_text = "\n".join([pointer_sign + pre_span_text, pointer_sign + curr_span_text, pointer_sign + next_span_text])
combined_span_texts.append(combined_text)
return "\n\n".join(combined_span_texts)
# pd.set_option("display.max_colwidth", None) # 设置为 None 来显示完整的文本
pd.set_option("display.max_rows", None) # 设置为 None 来显示更多的行
def main():
pdf_dict_files = __find_pdf_dic_files()
# print(pdf_dict_files)
span_stats_calc = SpanStatsCalc()
for pdf_dict_file in pdf_dict_files:
print("-" * 100)
print_green_on_red(f"Processing {pdf_dict_file}")
with open(pdf_dict_file, "r", encoding="utf-8") as f:
pdf_dict = json.load(f)
raw_df = span_stats_calc.calc_stats_per_dict(pdf_dict)
save_path = pdf_dict_file.replace("pdf_dic.json", "span_stats_raw.csv")
raw_df.to_csv(save_path, index=False)
filtered_df = raw_df[raw_df["span_is_superscript"] == 1]
if filtered_df.empty:
print("No superscript span found!")
continue
filtered_grouped_df = filtered_df.groupby(["span_font_name", "span_font_size", "span_font_color"])
combined_span_texts = filtered_grouped_df.apply(combine_span_texts, span_stats=raw_df) # type: ignore
final_df = filtered_grouped_df.size().reset_index(name="count")
final_df["span_texts"] = combined_span_texts.reset_index(level=[0, 1, 2], drop=True)
print(final_df)
final_df["span_texts"] = final_df["span_texts"].apply(lambda x: x.replace("\n", "\r\n"))
save_path = pdf_dict_file.replace("pdf_dic.json", "span_stats_final.csv")
# 使用 UTF-8 编码并添加 BOM，确保所有字段被双引号包围
final_df.to_csv(save_path, index=False, encoding="utf-8-sig", quoting=csv.QUOTE_ALL)
# 创建一个 2x2 的图表布局
fig, axs = plt.subplots(2, 2, figsize=(15, 10))
# 按照 span_font_name 分类作图
final_df.groupby("span_font_name")["count"].sum().plot(kind="bar", ax=axs[0, 0], title="By Font Name")
# 按照 span_font_size 分类作图
final_df.groupby("span_font_size")["count"].sum().plot(kind="bar", ax=axs[0, 1], title="By Font Size")
# 按照 span_font_color 分类作图
final_df.groupby("span_font_color")["count"].sum().plot(kind="bar", ax=axs[1, 0], title="By Font Color")
# 按照 span_font_name、span_font_size 和 span_font_color 共同分类作图
grouped = final_df.groupby(["span_font_name", "span_font_size", "span_font_color"])
grouped["count"].sum().unstack().plot(kind="bar", ax=axs[1, 1], title="Combined Grouping")
# 调整布局
plt.tight_layout()
# 显示图表
# plt.show()
# 保存图表到 PNG 文件
save_path = pdf_dict_file.replace("pdf_dic.json", "span_stats_combined.png")
plt.savefig(save_path)
# 清除画布
plt.clf()
if __name__ == "__main__":
main()
magic_pdf/libs/commons.py
View file @ 8afff9ae
import datetime
import json
import os, re, configparser
import subprocess
import time
import boto3
from loguru import logger
from boto3.s3.transfer import TransferConfig
from botocore.config import Config
import fitz # 1.23.9中已经切换到rebase
# import fitz_old as fitz # 使用1.23.9之前的pymupdf库
def get_delta_time(input_time):
return round(time.time() - input_time, 2)
def join_path(*args): def join_path(*args):
return '/'.join(str(s).rstrip('/') for s in args) return '/'.join(str(s).rstrip('/') for s in args)
#配置全局的errlog_path，方便demo同步引用
error_log_path = "s3://llm-pdf-text/err_logs/"
# json_dump_path = "s3://pdf_books_temp/json_dump/" # 这条路径仅用于临时本地测试,不能提交到main
json_dump_path = "s3://llm-pdf-text/json_dump/"
# s3_image_save_path = "s3://mllm-raw-media/pdf2md_img/" # 基础库不应该有这些存在的路径，应该在业务代码中定义
def get_top_percent_list(num_list, percent): def get_top_percent_list(num_list, percent):
""" """
获取列表中前百分之多少的元素 获取列表中前百分之多少的元素
...@@ -48,51 +22,12 @@ def get_top_percent_list(num_list, percent): ...@@ -48,51 +22,12 @@ def get_top_percent_list(num_list, percent):
return top_percent_list return top_percent_list
def formatted_time(time_stamp):
dt_object = datetime.datetime.fromtimestamp(time_stamp)
output_time = dt_object.strftime("%Y-%m-%d-%H:%M:%S")
return output_time
def mymax(alist: list): def mymax(alist: list):
if len(alist) == 0: if len(alist) == 0:
return 0 # 空是0， 0*0也是0大小q return 0 # 空是0， 0*0也是0大小q
else: else:
return max(alist) return max(alist)
def parse_aws_param(profile):
if isinstance(profile, str):
# 解析配置文件
config_file = join_path(os.path.expanduser("~"), ".aws", "config")
credentials_file = join_path(os.path.expanduser("~"), ".aws", "credentials")
config = configparser.ConfigParser()
config.read(credentials_file)
config.read(config_file)
# 获取 AWS 账户相关信息
ak = config.get(profile, "aws_access_key_id")
sk = config.get(profile, "aws_secret_access_key")
if profile == "default":
s3_str = config.get(f"{profile}", "s3")
else:
s3_str = config.get(f"profile {profile}", "s3")
end_match = re.search("endpoint_url[\s]*=[\s]*([^\s\n]+)[\s\n]*$", s3_str, re.MULTILINE)
if end_match:
endpoint = end_match.group(1)
else:
raise ValueError(f"aws 配置文件中没有找到 endpoint_url")
style_match = re.search("addressing_style[\s]*=[\s]*([^\s\n]+)[\s\n]*$", s3_str, re.MULTILINE)
if style_match:
addressing_style = style_match.group(1)
else:
addressing_style = "path"
elif isinstance(profile, dict):
ak = profile["ak"]
sk = profile["sk"]
endpoint = profile["endpoint"]
addressing_style = "auto"
return ak, sk, endpoint, addressing_style
def parse_bucket_key(s3_full_path: str): def parse_bucket_key(s3_full_path: str):
""" """
...@@ -106,99 +41,3 @@ def parse_bucket_key(s3_full_path: str): ...@@ -106,99 +41,3 @@ def parse_bucket_key(s3_full_path: str):
s3_full_path = s3_full_path[1:] s3_full_path = s3_full_path[1:]
bucket, key = s3_full_path.split("/", 1) bucket, key = s3_full_path.split("/", 1)
return bucket, key return bucket, key
def read_file(pdf_path: str, s3_profile):
if pdf_path.startswith("s3://"):
ak, sk, end_point, addressing_style = parse_aws_param(s3_profile)
cli = boto3.client(service_name="s3", aws_access_key_id=ak, aws_secret_access_key=sk, endpoint_url=end_point,
config=Config(s3={'addressing_style': addressing_style}, retries={'max_attempts': 10, 'mode': 'standard'}))
bucket_name, bucket_key = parse_bucket_key(pdf_path)
res = cli.get_object(Bucket=bucket_name, Key=bucket_key)
file_content = res["Body"].read()
return file_content
else:
with open(pdf_path, "rb") as f:
return f.read()
def get_docx_model_output(pdf_model_output, page_id):
model_output_json = pdf_model_output[page_id]
return model_output_json
def list_dir(dir_path:str, s3_profile:str):
"""
列出dir_path下的所有文件
"""
ret = []
if dir_path.startswith("s3"):
ak, sk, end_point, addressing_style = parse_aws_param(s3_profile)
s3info = re.findall(r"s3:\/\/([^\/]+)\/(.*)", dir_path)
bucket, path = s3info[0][0], s3info[0][1]
try:
cli = boto3.client(service_name="s3", aws_access_key_id=ak, aws_secret_access_key=sk, endpoint_url=end_point,
config=Config(s3={'addressing_style': addressing_style}))
def list_obj_scluster():
marker = None
while True:
list_kwargs = dict(MaxKeys=1000, Bucket=bucket, Prefix=path)
if marker:
list_kwargs['Marker'] = marker
response = cli.list_objects(**list_kwargs)
contents = response.get("Contents", [])
yield from contents
if not response.get("IsTruncated") or len(contents)==0:
break
marker = contents[-1]['Key']
for info in list_obj_scluster():
file_path = info['Key']
#size = info['Size']
if path!="":
afile = file_path[len(path):]
if afile.endswith(".json"):
ret.append(f"s3://{bucket}/{file_path}")
return ret
except Exception as e:
logger.exception(e)
exit(-1)
else: #本地的目录，那么扫描本地目录并返会这个目录里的所有jsonl文件
for root, dirs, files in os.walk(dir_path):
for file in files:
if file.endswith(".json"):
ret.append(join_path(root, file))
ret.sort()
return ret
def get_img_s3_client(save_path:str, image_s3_config:str):
"""
"""
if save_path.startswith("s3://"): # 放这里是为了最少创建一个s3 client
ak, sk, end_point, addressing_style = parse_aws_param(image_s3_config)
img_s3_client = boto3.client(
service_name="s3",
aws_access_key_id=ak,
aws_secret_access_key=sk,
endpoint_url=end_point,
config=Config(s3={"addressing_style": addressing_style}, retries={'max_attempts': 5, 'mode': 'standard'}),
)
else:
img_s3_client = None
return img_s3_client
if __name__=="__main__":
s3_path = "s3://llm-pdf-text/layout_det/scihub/scimag07865000-07865999/10.1007/s10729-011-9175-6.pdf/"
s3_profile = "langchao"
ret = list_dir(s3_path, s3_profile)
print(ret)
\ No newline at end of file
magic_pdf/libs/detect_language_from_model.py deleted 100644 → 0
View file @ 4df1eb74
from collections import Counter
from magic_pdf.libs.language import detect_lang
def get_language_from_model(model_list: list):
language_lst = []
for ocr_page_info in model_list:
page_text = ""
layout_dets = ocr_page_info["layout_dets"]
for layout_det in layout_dets:
category_id = layout_det["category_id"]
allow_category_id_list = [15]
if category_id in allow_category_id_list:
page_text += layout_det["text"]
page_language = detect_lang(page_text)
language_lst.append(page_language)
# 统计text_language_list中每种语言的个数
count_dict = Counter(language_lst)
# 输出text_language_list中出现的次数最多的语言
language = max(count_dict, key=count_dict.get)
return language
magic_pdf/libs/draw_bbox.py
View file @ 8afff9ae
import fitz
from magic_pdf.config.constants import CROSS_PAGE from magic_pdf.config.constants import CROSS_PAGE
from magic_pdf.config.ocr_content_type import (BlockType, CategoryId, from magic_pdf.config.ocr_content_type import BlockType, CategoryId, ContentType
ContentType)
from magic_pdf.data.dataset import PymuDocDataset from magic_pdf.data.dataset import PymuDocDataset
from magic_pdf.libs.commons import fitz # PyMuPDF
from magic_pdf.model.magic_model import MagicModel from magic_pdf.model.magic_model import MagicModel
......
magic_pdf/libs/markdown_utils.py
View file @ 8afff9ae
import re
def escape_special_markdown_char(pymu_blocks):
"""
转义正文里对markdown语法有特殊意义的字符
"""
special_chars = ["*", "`", "~", "$"]
for blk in pymu_blocks:
for line in blk['lines']:
for span in line['spans']:
for char in special_chars:
span_text = span['text']
span_type = span.get("_type", None)
if span_type in ['inline-equation', 'interline-equation']:
continue
elif span_text:
span['text'] = span['text'].replace(char, "\\" + char)
return pymu_blocks
def ocr_escape_special_markdown_char(content): def ocr_escape_special_markdown_char(content):
""" """
......
magic_pdf/libs/nlp_utils.py deleted 100644 → 0
View file @ 4df1eb74
import re
from os import path
from collections import Counter
from loguru import logger
# from langdetect import detect
import spacy
import en_core_web_sm
import zh_core_web_sm
from magic_pdf.libs.language import detect_lang
class NLPModels:
"""
How to upload local models to s3:
- config aws cli:
doc\SETUP-CLI.md
doc\setup_cli.sh
app\config\__init__.py
- $ cd {local_dir_storing_models}
- $ ls models
en_core_web_sm-3.7.1/
zh_core_web_sm-3.7.0/
- $ aws s3 sync models/ s3://llm-infra/models --profile=p_project_norm
- $ aws s3 --profile=p_project_norm ls s3://llm-infra/models/
PRE en_core_web_sm-3.7.1/
PRE zh_core_web_sm-3.7.0/
"""
def __init__(self):
# if OS is windows, set "TMP_DIR" to "D:/tmp"
home_dir = path.expanduser("~")
self.default_local_path = path.join(home_dir, ".nlp_models")
self.default_shared_path = "/share/pdf_processor/nlp_models"
self.default_hdfs_path = "hdfs://pdf_processor/nlp_models"
self.default_s3_path = "s3://llm-infra/models"
self.nlp_models = self.nlp_models = {
"en_core_web_sm": {
"type": "spacy",
"version": "3.7.1",
},
"en_core_web_md": {
"type": "spacy",
"version": "3.7.1",
},
"en_core_web_lg": {
"type": "spacy",
"version": "3.7.1",
},
"zh_core_web_sm": {
"type": "spacy",
"version": "3.7.0",
},
"zh_core_web_md": {
"type": "spacy",
"version": "3.7.0",
},
"zh_core_web_lg": {
"type": "spacy",
"version": "3.7.0",
},
}
self.en_core_web_sm_model = en_core_web_sm.load()
self.zh_core_web_sm_model = zh_core_web_sm.load()
def load_model(self, model_name, model_type, model_version):
if (
model_name in self.nlp_models
and self.nlp_models[model_name]["type"] == model_type
and self.nlp_models[model_name]["version"] == model_version
):
return spacy.load(model_name) if spacy.util.is_package(model_name) else None
else:
logger.error(f"Unsupported model name or version: {model_name} {model_version}")
return None
def detect_language(self, text, use_langdetect=False):
if len(text) == 0:
return None
if use_langdetect:
# print("use_langdetect")
# print(detect_lang(text))
# return detect_lang(text)
if detect_lang(text) == "zh":
return "zh"
else:
return "en"
if not use_langdetect:
en_count = len(re.findall(r"[a-zA-Z]", text))
cn_count = len(re.findall(r"[\u4e00-\u9fff]", text))
if en_count > cn_count:
return "en"
if cn_count > en_count:
return "zh"
def detect_entity_catgr_using_nlp(self, text, threshold=0.5):
"""
Detect entity categories using NLP models and return the most frequent entity types.
Parameters
----------
text : str
Text to be processed.
Returns
-------
str
The most frequent entity type.
"""
lang = self.detect_language(text, use_langdetect=True)
if lang == "en":
nlp_model = self.en_core_web_sm_model
elif lang == "zh":
nlp_model = self.zh_core_web_sm_model
else:
# logger.error(f"Unsupported language: {lang}")
return {}
# Splitting text into smaller parts
text_parts = re.split(r"[,;，；、\s & |]+", text)
text_parts = [part for part in text_parts if not re.match(r"[\d\W]+", part)] # Remove non-words
text_combined = " ".join(text_parts)
try:
doc = nlp_model(text_combined)
entity_counts = Counter([ent.label_ for ent in doc.ents])
word_counts_in_entities = Counter()
for ent in doc.ents:
word_counts_in_entities[ent.label_] += len(ent.text.split())
total_words_in_entities = sum(word_counts_in_entities.values())
total_words = len([token for token in doc if not token.is_punct])
if total_words_in_entities == 0 or total_words == 0:
return None
entity_percentage = total_words_in_entities / total_words
if entity_percentage < 0.5:
return None
most_common_entity, word_count = word_counts_in_entities.most_common(1)[0]
entity_percentage = word_count / total_words_in_entities
if entity_percentage >= threshold:
return most_common_entity
else:
return None
except Exception as e:
logger.error(f"Error in entity detection: {e}")
return None
def __main__():
nlpModel = NLPModels()
test_strings = [
"张三",
"张三, 李四，王五; 赵六",
"John Doe",
"Jane Smith",
"Lee, John",
"John Doe, Jane Smith; Alice Johnson，Bob Lee",
"孙七, Michael Jordan；赵八",
"David Smith Michael O'Connor; Kevin ßáçøñ",
"李雷·韩梅梅, 张三·李四",
"Charles Robert Darwin, Isaac Newton",
"莱昂纳多·迪卡普里奥, 杰克·吉伦哈尔",
"John Doe, Jane Smith; Alice Johnson",
"张三, 李四，王五; 赵六",
"Lei Wang, Jia Li, and Xiaojun Chen, LINKE YANG OU, and YUAN ZHANG",
"Rachel Mills & William Barry & Susanne B. Haga",
"Claire Chabut* and Jean-François Bussières",
"1 Department of Chemistry, Northeastern University, Shenyang 110004, China 2 State Key Laboratory of Polymer Physics and Chemistry, Changchun Institute of Applied Chemistry, Chinese Academy of Sciences, Changchun 130022, China",
"Changchun",
"china",
"Rongjun Song, 1,2 Baoyan Zhang, 1 Baotong Huang, 2 Tao Tang 2",
"Synergistic Effect of Supported Nickel Catalyst with Intumescent Flame-Retardants on Flame Retardancy and Thermal Stability of Polypropylene",
"Synergistic Effect of Supported Nickel Catalyst with",
"Intumescent Flame-Retardants on Flame Retardancy",
"and Thermal Stability of Polypropylene",
]
for test in test_strings:
print()
print(f"Original String: {test}")
result = nlpModel.detect_entity_catgr_using_nlp(test)
print(f"Detected entities: {result}")
if __name__ == "__main__":
__main__()
magic_pdf/libs/pdf_image_tools.py
View file @ 8afff9ae
from io import BytesIO from io import BytesIO
import cv2 import cv2
import fitz
import numpy as np import numpy as np
from PIL import Image from PIL import Image
from magic_pdf.data.data_reader_writer import DataWriter from magic_pdf.data.data_reader_writer import DataWriter
from magic_pdf.libs.commons import fitz, join_path from magic_pdf.libs.commons import join_path
from magic_pdf.libs.hash_utils import compute_sha256 from magic_pdf.libs.hash_utils import compute_sha256
......
magic_pdf/libs/textbase.py deleted 100644 → 0
View file @ 4df1eb74
import math
def __inc_dict_val(mp, key, val_inc:int):
if mp.get(key):
mp[key] = mp[key] + val_inc
else:
mp[key] = val_inc
def get_text_block_base_info(block):
"""
获取这个文本块里的字体的颜色、字号、字体
按照正文字数最多的返回
"""
counter = {}
for line in block['lines']:
for span in line['spans']:
color = span['color']
size = round(span['size'], 2)
font = span['font']
txt_len = len(span['text'])
__inc_dict_val(counter, (color, size, font), txt_len)
c, s, ft = max(counter, key=counter.get)
return c, s, ft
\ No newline at end of file
magic_pdf/libs/vis_utils.py deleted 100644 → 0
View file @ 4df1eb74
from magic_pdf.libs.commons import fitz
import os
def draw_bbox_on_page(raw_pdf_doc: fitz.Document, paras_dict:dict, save_path: str):
"""
在page上画出bbox，保存到save_path
"""
# 检查文件是否存在
is_new_pdf = False
if os.path.exists(save_path):
# 打开现有的 PDF 文件
doc = fitz.open(save_path)
else:
# 创建一个新的空白 PDF 文件
is_new_pdf = True
doc = fitz.open('')
color_map = {
'image': fitz.pdfcolor["yellow"],
'text': fitz.pdfcolor['blue'],
"table": fitz.pdfcolor['green']
}
for k, v in paras_dict.items():
page_idx = v['page_idx']
width = raw_pdf_doc[page_idx].rect.width
height = raw_pdf_doc[page_idx].rect.height
new_page = doc.new_page(width=width, height=height)
shape = new_page.new_shape()
for order, block in enumerate(v['preproc_blocks']):
rect = fitz.Rect(block['bbox'])
shape = new_page.new_shape()
shape.draw_rect(rect)
shape.finish(color=None, fill=color_map['text'], fill_opacity=0.2)
shape.finish()
shape.commit()
for img in v['images']:
# 原始box画上去
rect = fitz.Rect(img['bbox'])
shape = new_page.new_shape()
shape.draw_rect(rect)
shape.finish(color=None, fill=fitz.pdfcolor['yellow'])
shape.finish()
shape.commit()
for img in v['image_backup']:
# 原始box画上去
rect = fitz.Rect(img['bbox'])
shape = new_page.new_shape()
shape.draw_rect(rect)
shape.finish(color=fitz.pdfcolor['yellow'], fill=None)
shape.finish()
shape.commit()
for tb in v['droped_text_block']:
# 原始box画上去
rect = fitz.Rect(tb['bbox'])
shape = new_page.new_shape()
shape.draw_rect(rect)
shape.finish(color=None, fill=fitz.pdfcolor['black'], fill_opacity=0.4)
shape.finish()
shape.commit()
# TODO table
for tb in v['tables']:
rect = fitz.Rect(tb['bbox'])
shape = new_page.new_shape()
shape.draw_rect(rect)
shape.finish(color=None, fill=fitz.pdfcolor['green'], fill_opacity=0.2)
shape.finish()
shape.commit()
parent_dir = os.path.dirname(save_path)
if not os.path.exists(parent_dir):
os.makedirs(parent_dir)
if is_new_pdf:
doc.save(save_path)
else:
doc.saveIncr()
doc.close()
def debug_show_bbox(raw_pdf_doc: fitz.Document, page_idx: int, bboxes: list, droped_bboxes:list, expect_drop_bboxes:list, save_path: str, expected_page_id:int):
"""
以覆盖的方式写个临时的pdf，用于debug
"""
if page_idx!=expected_page_id:
return
if os.path.exists(save_path):
# 删除已经存在的文件
os.remove(save_path)
# 创建一个新的空白 PDF 文件
doc = fitz.open('')
width = raw_pdf_doc[page_idx].rect.width
height = raw_pdf_doc[page_idx].rect.height
new_page = doc.new_page(width=width, height=height)
shape = new_page.new_shape()
for bbox in bboxes:
# 原始box画上去
rect = fitz.Rect(*bbox[0:4])
shape = new_page.new_shape()
shape.draw_rect(rect)
shape.finish(color=fitz.pdfcolor['red'], fill=fitz.pdfcolor['blue'], fill_opacity=0.2)
shape.finish()
shape.commit()
for bbox in droped_bboxes:
# 原始box画上去
rect = fitz.Rect(*bbox[0:4])
shape = new_page.new_shape()
shape.draw_rect(rect)
shape.finish(color=None, fill=fitz.pdfcolor['yellow'], fill_opacity=0.2)
shape.finish()
shape.commit()
for bbox in expect_drop_bboxes:
# 原始box画上去
rect = fitz.Rect(*bbox[0:4])
shape = new_page.new_shape()
shape.draw_rect(rect)
shape.finish(color=fitz.pdfcolor['red'], fill=None)
shape.finish()
shape.commit()
# shape.insert_textbox(fitz.Rect(200, 0, 600, 20), f"total bboxes: {len(bboxes)}", fontname="helv", fontsize=12,
# color=(0, 0, 0))
# shape.finish(color=fitz.pdfcolor['black'])
# shape.commit()
parent_dir = os.path.dirname(save_path)
if not os.path.exists(parent_dir):
os.makedirs(parent_dir)
doc.save(save_path)
doc.close()
def debug_show_page(page, bboxes1: list,bboxes2: list,bboxes3: list,):
save_path = "./tmp/debug.pdf"
if os.path.exists(save_path):
# 删除已经存在的文件
os.remove(save_path)
# 创建一个新的空白 PDF 文件
doc = fitz.open('')
width = page.rect.width
height = page.rect.height
new_page = doc.new_page(width=width, height=height)
shape = new_page.new_shape()
for bbox in bboxes1:
# 原始box画上去
rect = fitz.Rect(*bbox[0:4])
shape = new_page.new_shape()
shape.draw_rect(rect)
shape.finish(color=fitz.pdfcolor['red'], fill=fitz.pdfcolor['blue'], fill_opacity=0.2)
shape.finish()
shape.commit()
for bbox in bboxes2:
# 原始box画上去
rect = fitz.Rect(*bbox[0:4])
shape = new_page.new_shape()
shape.draw_rect(rect)
shape.finish(color=None, fill=fitz.pdfcolor['yellow'], fill_opacity=0.2)
shape.finish()
shape.commit()
for bbox in bboxes3:
# 原始box画上去
rect = fitz.Rect(*bbox[0:4])
shape = new_page.new_shape()
shape.draw_rect(rect)
shape.finish(color=fitz.pdfcolor['red'], fill=None)
shape.finish()
shape.commit()
parent_dir = os.path.dirname(save_path)
if not os.path.exists(parent_dir):
os.makedirs(parent_dir)
doc.save(save_path)
doc.close()
def draw_layout_bbox_on_page(raw_pdf_doc: fitz.Document, paras_dict:dict, header, footer, pdf_path: str):
"""
在page上画出bbox，保存到save_path
"""
# 检查文件是否存在
is_new_pdf = False
if os.path.exists(pdf_path):
# 打开现有的 PDF 文件
doc = fitz.open(pdf_path)
else:
# 创建一个新的空白 PDF 文件
is_new_pdf = True
doc = fitz.open('')
for k, v in paras_dict.items():
page_idx = v['page_idx']
layouts = v['layout_bboxes']
page = doc[page_idx]
shape = page.new_shape()
for order, layout in enumerate(layouts):
border_offset = 1
rect_box = layout['layout_bbox']
layout_label = layout['layout_label']
fill_color = fitz.pdfcolor['pink'] if layout_label=='U' else None
rect_box = [rect_box[0]+1, rect_box[1]-border_offset, rect_box[2]-1, rect_box[3]+border_offset]
rect = fitz.Rect(*rect_box)
shape.draw_rect(rect)
shape.finish(color=fitz.pdfcolor['red'], fill=fill_color, fill_opacity=0.4)
"""
draw order text on layout box
"""
font_size = 10
shape.insert_text((rect_box[0] + 1, rect_box[1] + font_size), f"{order}", fontsize=font_size, color=(0, 0, 0))
"""画上footer header"""
if header:
shape.draw_rect(fitz.Rect(header))
shape.finish(color=None, fill=fitz.pdfcolor['black'], fill_opacity=0.2)
if footer:
shape.draw_rect(fitz.Rect(footer))
shape.finish(color=None, fill=fitz.pdfcolor['black'], fill_opacity=0.2)
shape.commit()
if is_new_pdf:
doc.save(pdf_path)
else:
doc.saveIncr()
doc.close()
@DeprecationWarning
def draw_layout_on_page(raw_pdf_doc: fitz.Document, page_idx: int, page_layout: list, pdf_path: str):
"""
把layout的box用红色边框花在pdf_path的page_idx上
"""
def draw(shape, layout, fill_color=fitz.pdfcolor['pink']):
border_offset = 1
rect_box = layout['layout_bbox']
layout_label = layout['layout_label']
sub_layout = layout['sub_layout']
if len(sub_layout)==0:
fill_color = fill_color if layout_label=='U' else None
rect_box = [rect_box[0]+1, rect_box[1]-border_offset, rect_box[2]-1, rect_box[3]+border_offset]
rect = fitz.Rect(*rect_box)
shape.draw_rect(rect)
shape.finish(color=fitz.pdfcolor['red'], fill=fill_color, fill_opacity=0.2)
# if layout_label=='U':
# bad_boxes = layout.get("bad_boxes", [])
# for bad_box in bad_boxes:
# rect = fitz.Rect(*bad_box)
# shape.draw_rect(rect)
# shape.finish(color=fitz.pdfcolor['red'], fill=fitz.pdfcolor['red'], fill_opacity=0.2)
# else:
# rect = fitz.Rect(*rect_box)
# shape.draw_rect(rect)
# shape.finish(color=fitz.pdfcolor['blue'])
for sub_layout in sub_layout:
draw(shape, sub_layout)
shape.commit()
# 检查文件是否存在
is_new_pdf = False
if os.path.exists(pdf_path):
# 打开现有的 PDF 文件
doc = fitz.open(pdf_path)
else:
# 创建一个新的空白 PDF 文件
is_new_pdf = True
doc = fitz.open('')
page = doc[page_idx]
shape = page.new_shape()
for order, layout in enumerate(page_layout):
draw(shape, layout, fitz.pdfcolor['yellow'])
# shape.insert_textbox(fitz.Rect(200, 0, 600, 20), f"total bboxes: {len(layout)}", fontname="helv", fontsize=12,
# color=(0, 0, 0))
# shape.finish(color=fitz.pdfcolor['black'])
# shape.commit()
parent_dir = os.path.dirname(pdf_path)
if not os.path.exists(parent_dir):
os.makedirs(parent_dir)
if is_new_pdf:
doc.save(pdf_path)
else:
doc.saveIncr()
doc.close()
\ No newline at end of file
magic_pdf/model/doc_analyze_by_custom_model.py
View file @ 8afff9ae
...@@ -46,8 +46,8 @@ def load_images_from_pdf(pdf_bytes: bytes, dpi=200, start_page_id=0, end_page_id ...@@ -46,8 +46,8 @@ def load_images_from_pdf(pdf_bytes: bytes, dpi=200, start_page_id=0, end_page_id
mat = fitz.Matrix(dpi / 72, dpi / 72) mat = fitz.Matrix(dpi / 72, dpi / 72)
pm = page.get_pixmap(matrix=mat, alpha=False) pm = page.get_pixmap(matrix=mat, alpha=False)
# If the width or height exceeds 9000 after scaling, do not scale further. # If the width or height exceeds 4500 after scaling, do not scale further.
if pm.width > 9000 or pm.height > 9000: if pm.width > 4500 or pm.height > 4500:
pm = page.get_pixmap(matrix=fitz.Matrix(1, 1), alpha=False) pm = page.get_pixmap(matrix=fitz.Matrix(1, 1), alpha=False)
img = Image.frombytes("RGB", (pm.width, pm.height), pm.samples) img = Image.frombytes("RGB", (pm.width, pm.height), pm.samples)
......
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