Merge pull request #2625 from opendatalab/release-2.0.0

Release 2.0.0

mineru/utils/format_utils.py

+
+import re
+import itertools
+import html
+from typing import Any, Dict, List
+from pydantic import (
+    BaseModel,
+    computed_field,
+    model_validator,
+)
+
+class TableCell(BaseModel):
+    """TableCell."""
+    row_span: int = 1
+    col_span: int = 1
+    start_row_offset_idx: int
+    end_row_offset_idx: int
+    start_col_offset_idx: int
+    end_col_offset_idx: int
+    text: str
+    column_header: bool = False
+    row_header: bool = False
+    row_section: bool = False
+
+    @model_validator(mode="before")
+    @classmethod
+    def from_dict_format(cls, data: Any) -> Any:
+        """from_dict_format."""
+        if isinstance(data, Dict):
+            # Check if this is a native BoundingBox or a bbox from docling-ibm-models
+            if (
+                    # "bbox" not in data
+                    # or data["bbox"] is None
+                    # or isinstance(data["bbox"], BoundingBox)
+                    "text"
+                    in data
+            ):
+                return data
+            text = data["bbox"].get("token", "")
+            if not len(text):
+                text_cells = data.pop("text_cell_bboxes", None)
+                if text_cells:
+                    for el in text_cells:
+                        text += el["token"] + " "
+
+                text = text.strip()
+            data["text"] = text
+
+        return data
+
+
+class TableData(BaseModel):  # TBD
+    """BaseTableData."""
+
+    table_cells: List[TableCell] = []
+    num_rows: int = 0
+    num_cols: int = 0
+
+    @computed_field  # type: ignore
+    @property
+    def grid(
+            self,
+    ) -> List[List[TableCell]]:
+        """grid."""
+        # Initialise empty table data grid (only empty cells)
+        table_data = [
+            [
+                TableCell(
+                    text="",
+                    start_row_offset_idx=i,
+                    end_row_offset_idx=i + 1,
+                    start_col_offset_idx=j,
+                    end_col_offset_idx=j + 1,
+                )
+                for j in range(self.num_cols)
+            ]
+            for i in range(self.num_rows)
+        ]
+
+        # Overwrite cells in table data for which there is actual cell content.
+        for cell in self.table_cells:
+            for i in range(
+                    min(cell.start_row_offset_idx, self.num_rows),
+                    min(cell.end_row_offset_idx, self.num_rows),
+            ):
+                for j in range(
+                        min(cell.start_col_offset_idx, self.num_cols),
+                        min(cell.end_col_offset_idx, self.num_cols),
+                ):
+                    table_data[i][j] = cell
+
+        return table_data
+
+
+"""
+OTSL
+"""
+OTSL_NL = "<nl>"
+OTSL_FCEL = "<fcel>"
+OTSL_ECEL = "<ecel>"
+OTSL_LCEL = "<lcel>"
+OTSL_UCEL = "<ucel>"
+OTSL_XCEL = "<xcel>"
+
+
+def otsl_extract_tokens_and_text(s: str):
+    # Pattern to match anything enclosed by < >
+    # (including the angle brackets themselves)
+    # pattern = r"(<[^>]+>)"
+    pattern = r"(" + r"|".join([OTSL_NL, OTSL_FCEL, OTSL_ECEL, OTSL_LCEL, OTSL_UCEL, OTSL_XCEL]) + r")"
+    # Find all tokens (e.g. "<otsl>", "<loc_140>", etc.)
+    tokens = re.findall(pattern, s)
+    # Remove any tokens that start with "<loc_"
+    tokens = [token for token in tokens]
+    # Split the string by those tokens to get the in-between text
+    text_parts = re.split(pattern, s)
+    text_parts = [token for token in text_parts]
+    # Remove any empty or purely whitespace strings from text_parts
+    text_parts = [part for part in text_parts if part.strip()]
+
+    return tokens, text_parts
+
+
+def otsl_parse_texts(texts, tokens):
+    split_word = OTSL_NL
+    split_row_tokens = [
+        list(y)
+        for x, y in itertools.groupby(tokens, lambda z: z == split_word)
+        if not x
+    ]
+    table_cells = []
+    r_idx = 0
+    c_idx = 0
+
+    def count_right(tokens, c_idx, r_idx, which_tokens):
+        span = 0
+        c_idx_iter = c_idx
+        while tokens[r_idx][c_idx_iter] in which_tokens:
+            c_idx_iter += 1
+            span += 1
+            if c_idx_iter >= len(tokens[r_idx]):
+                return span
+        return span
+
+    def count_down(tokens, c_idx, r_idx, which_tokens):
+        span = 0
+        r_idx_iter = r_idx
+        while tokens[r_idx_iter][c_idx] in which_tokens:
+            r_idx_iter += 1
+            span += 1
+            if r_idx_iter >= len(tokens):
+                return span
+        return span
+
+    for i, text in enumerate(texts):
+        cell_text = ""
+        if text in [
+            OTSL_FCEL,
+            OTSL_ECEL,
+        ]:
+            row_span = 1
+            col_span = 1
+            right_offset = 1
+            if text != OTSL_ECEL:
+                cell_text = texts[i + 1]
+                right_offset = 2
+
+            # Check next element(s) for lcel / ucel / xcel,
+            # set properly row_span, col_span
+            next_right_cell = ""
+            if i + right_offset < len(texts):
+                next_right_cell = texts[i + right_offset]
+
+            next_bottom_cell = ""
+            if r_idx + 1 < len(split_row_tokens):
+                if c_idx < len(split_row_tokens[r_idx + 1]):
+                    next_bottom_cell = split_row_tokens[r_idx + 1][c_idx]
+
+            if next_right_cell in [
+                OTSL_LCEL,
+                OTSL_XCEL,
+            ]:
+                # we have horisontal spanning cell or 2d spanning cell
+                col_span += count_right(
+                    split_row_tokens,
+                    c_idx + 1,
+                    r_idx,
+                    [OTSL_LCEL, OTSL_XCEL],
+                )
+            if next_bottom_cell in [
+                OTSL_UCEL,
+                OTSL_XCEL,
+            ]:
+                # we have a vertical spanning cell or 2d spanning cell
+                row_span += count_down(
+                    split_row_tokens,
+                    c_idx,
+                    r_idx + 1,
+                    [OTSL_UCEL, OTSL_XCEL],
+                )
+
+            table_cells.append(
+                TableCell(
+                    text=cell_text.strip(),
+                    row_span=row_span,
+                    col_span=col_span,
+                    start_row_offset_idx=r_idx,
+                    end_row_offset_idx=r_idx + row_span,
+                    start_col_offset_idx=c_idx,
+                    end_col_offset_idx=c_idx + col_span,
+                )
+            )
+        if text in [
+            OTSL_FCEL,
+            OTSL_ECEL,
+            OTSL_LCEL,
+            OTSL_UCEL,
+            OTSL_XCEL,
+        ]:
+            c_idx += 1
+        if text == OTSL_NL:
+            r_idx += 1
+            c_idx = 0
+    return table_cells, split_row_tokens
+
+
+def export_to_html(table_data: TableData):
+    nrows = table_data.num_rows
+    ncols = table_data.num_cols
+
+    text = ""
+
+    if len(table_data.table_cells) == 0:
+        return ""
+
+    body = ""
+
+    for i in range(nrows):
+        body += "<tr>"
+        for j in range(ncols):
+            cell: TableCell = table_data.grid[i][j]
+
+            rowspan, rowstart = (
+                cell.row_span,
+                cell.start_row_offset_idx,
+            )
+            colspan, colstart = (
+                cell.col_span,
+                cell.start_col_offset_idx,
+            )
+
+            if rowstart != i:
+                continue
+            if colstart != j:
+                continue
+
+            content = html.escape(cell.text.strip())
+            celltag = "td"
+            if cell.column_header:
+                celltag = "th"
+
+            opening_tag = f"{celltag}"
+            if rowspan > 1:
+                opening_tag += f' rowspan="{rowspan}"'
+            if colspan > 1:
+                opening_tag += f' colspan="{colspan}"'
+
+            body += f"<{opening_tag}>{content}</{celltag}>"
+        body += "</tr>"
+
+    # dir = get_text_direction(text)
+    body = f"<table>{body}</table>"
+
+    return body
+
+
+def convert_otsl_to_html(otsl_content: str):
+    tokens, mixed_texts = otsl_extract_tokens_and_text(otsl_content)
+    table_cells, split_row_tokens = otsl_parse_texts(mixed_texts, tokens)
+
+    table_data = TableData(
+        num_rows=len(split_row_tokens),
+        num_cols=(
+            max(len(row) for row in split_row_tokens) if split_row_tokens else 0
+        ),
+        table_cells=table_cells,
+    )
+
+    return export_to_html(table_data)

magic_pdf/libs/hash_utils.py → mineru/utils/hash_utils.py

+# Copyright (c) Opendatalab. All rights reserved.
 import hashlib
+import json
 
 
-def compute_md5(file_bytes):
+def bytes_md5(file_bytes):
     hasher = hashlib.md5()
     hasher.update(file_bytes)
     return hasher.hexdigest().upper()
 
 
-def compute_sha256(input_string):
+def str_md5(input_string):
+    hasher = hashlib.md5()
+    # 在Python3中，需要将字符串转化为字节对象才能被哈希函数处理
+    input_bytes = input_string.encode('utf-8')
+    hasher.update(input_bytes)
+    return hasher.hexdigest()
+
+
+def str_sha256(input_string):
     hasher = hashlib.sha256()
     # 在Python3中，需要将字符串转化为字节对象才能被哈希函数处理
     input_bytes = input_string.encode('utf-8')
     hasher.update(input_bytes)
     return hasher.hexdigest()
+
+
+def dict_md5(d):
+    json_str = json.dumps(d, sort_keys=True, ensure_ascii=False)
+    return hashlib.md5(json_str.encode('utf-8')).hexdigest()
\ No newline at end of file

magic_pdf/post_proc/llm_aided.py → mineru/utils/llm_aided.py

 # Copyright (c) Opendatalab. All rights reserved.
-import json
 from loguru import logger
-from magic_pdf.dict2md.ocr_mkcontent import merge_para_with_text
 from openai import OpenAI
 import ast
 
+from mineru.backend.pipeline.pipeline_middle_json_mkcontent import merge_para_with_text
 
-#@todo: 有的公式以"\"结尾，这样会导致尾部拼接的"$"被转义，也需要修复
-formula_optimize_prompt = """请根据以下指南修正LaTeX公式的错误，确保公式能够渲染且符合原始内容：
 
-1. 修正渲染或编译错误：
-    - Some syntax errors such as mismatched/missing/extra tokens. Your task is to fix these syntax errors and make sure corrected results conform to latex math syntax principles.
-    - 包含KaTeX不支持的关键词等原因导致的无法编译或渲染的错误
-
-2. 保留原始信息：
-   - 保留原始公式中的所有重要信息
-   - 不要添加任何原始公式中没有的新信息
-
-IMPORTANT:请仅返回修正后的公式，不要包含任何介绍、解释或元数据。
-
-LaTeX recognition result:
-$FORMULA
-
-Your corrected result:
-"""
-
-text_optimize_prompt = f"""请根据以下指南修正OCR引起的错误，确保文本连贯并符合原始内容：
-
-1. 修正OCR引起的拼写错误和错误：
-   - 修正常见的OCR错误（例如，'rn' 被误读为 'm'）
-   - 使用上下文和常识进行修正
-   - 只修正明显的错误，不要不必要的修改内容
-   - 不要添加额外的句号或其他不必要的标点符号
-
-2. 保持原始结构：
-   - 保留所有标题和子标题
-
-3. 保留原始内容：
-   - 保留原始文本中的所有重要信息
-   - 不要添加任何原始文本中没有的新信息
-   - 保留段落之间的换行符
-
-4. 保持连贯性：
-   - 确保内容与前文顺畅连接
-   - 适当处理在句子中间开始或结束的文本
-   
-5. 修正行内公式：
-   - 去除行内公式前后多余的空格
-   - 修正公式中的OCR错误
-   - 确保公式能够通过KaTeX渲染
-   
-6. 修正全角字符
-    - 修正全角标点符号为半角标点符号
-    - 修正全角字母为半角字母
-    - 修正全角数字为半角数字
-
-IMPORTANT:请仅返回修正后的文本，保留所有原始格式，包括换行符。不要包含任何介绍、解释或元数据。
-
-Previous context:
-
-Current chunk to process:
-
-Corrected text:
-"""
-
-def llm_aided_formula(pdf_info_dict, formula_aided_config):
-    pass
-
-def llm_aided_text(pdf_info_dict, text_aided_config):
-    pass
-
-def llm_aided_title(pdf_info_dict, title_aided_config):
+def llm_aided_title(page_info_list, title_aided_config):
     client = OpenAI(
         api_key=title_aided_config["api_key"],
         base_url=title_aided_config["base_url"],
@@ -78,8 +14,8 @@ def llm_aided_title(pdf_info_dict, title_aided_config):
     title_dict = {}
     origin_title_list = []
     i = 0
-    for page_num, page in pdf_info_dict.items():
-        blocks = page["para_blocks"]
+    for page_info in page_info_list:
+        blocks = page_info["para_blocks"]
         for block in blocks:
             if block["type"] == "title":
                 origin_title_list.append(block)
@@ -92,7 +28,7 @@ def llm_aided_title(pdf_info_dict, title_aided_config):
                     line_avg_height = sum(page_line_height_list) / len(page_line_height_list)
                 else:
                     line_avg_height = int(block['bbox'][3] - block['bbox'][1])
-                title_dict[f"{i}"] = [title_text, line_avg_height, int(page_num[5:])+1]
+                title_dict[f"{i}"] = [title_text, line_avg_height, int(page_info['page_idx']) + 1]
                 i += 1
     # logger.info(f"Title list: {title_dict}")
 
@@ -115,16 +51,21 @@ def llm_aided_title(pdf_info_dict, title_aided_config):
     - 标题从前至后的层级必须是连续的，不能跳过层级
     - 标题层级最多为4级，不要添加过多的层级
     - 优化后的标题只保留代表该标题的层级的整数，不要保留其他信息
-    
+
 5. 合理性检查与微调：
     - 在完成初步分级后，仔细检查分级结果的合理性
     - 根据上下文关系和逻辑顺序，对不合理的分级进行微调
     - 确保最终的分级结果符合文档的实际结构和逻辑
     - 字典中可能包含被误当成标题的正文，你可以通过将其层级标记为 0 来排除它们
-    
+
 IMPORTANT: 
 请直接返回优化过的由标题层级组成的字典，格式为{{标题id:标题层级}}，如下：
-{{0:1,1:2,2:2,3:3}}
+{{
+  0:1,
+  1:2,
+  2:2,
+  3:3
+}}
 不需要对字典格式化，不需要返回任何其他信息。
 
 Input title list:
@@ -145,16 +86,23 @@ Corrected title list:
                     {'role': 'user', 'content': title_optimize_prompt}],
                 temperature=0.7,
             )
-            # logger.info(f"Title completion: {completion.choices[0].message.content}")
-            dict_completion = ast.literal_eval(completion.choices[0].message.content)
-            # logger.info(f"len(dict_completion): {len(dict_completion)}, len(title_dict): {len(title_dict)}")
+            content = completion.choices[0].message.content.strip()
+            # logger.info(f"Title completion: {content}")
+            if "</think>" in content:
+                idx = content.index("</think>") + len("</think>")
+                content = content[idx:].strip()
+            import json_repair
+            dict_completion = json_repair.loads(content)
+            dict_completion = {int(k): int(v) for k, v in dict_completion.items()}
 
+            # logger.info(f"len(dict_completion): {len(dict_completion)}, len(title_dict): {len(title_dict)}")
             if len(dict_completion) == len(title_dict):
                 for i, origin_title_block in enumerate(origin_title_list):
                     origin_title_block["level"] = int(dict_completion[i])
                 break
             else:
-                logger.warning("The number of titles in the optimized result is not equal to the number of titles in the input.")
+                logger.warning(
+                    "The number of titles in the optimized result is not equal to the number of titles in the input.")
                 retry_count += 1
         except Exception as e:
             logger.exception(e)

magic_pdf/model/sub_modules/model_utils.py → mineru/utils/model_utils.py

 import time
 import torch
+import gc
+from PIL import Image
 from loguru import logger
 import numpy as np
 
-from magic_pdf.libs.boxbase import get_minbox_if_overlap_by_ratio
-from magic_pdf.libs.clean_memory import clean_memory
+from mineru.utils.boxbase import get_minbox_if_overlap_by_ratio
 
 
-def crop_img(input_res, input_np_img, crop_paste_x=0, crop_paste_y=0):
+def crop_img(input_res, input_img, crop_paste_x=0, crop_paste_y=0):
 
     crop_xmin, crop_ymin = int(input_res['poly'][0]), int(input_res['poly'][1])
     crop_xmax, crop_ymax = int(input_res['poly'][4]), int(input_res['poly'][5])
@@ -16,15 +17,24 @@ def crop_img(input_res, input_np_img, crop_paste_x=0, crop_paste_y=0):
     crop_new_width = crop_xmax - crop_xmin + crop_paste_x * 2
     crop_new_height = crop_ymax - crop_ymin + crop_paste_y * 2
 
-    # Create a white background array
-    return_image = np.ones((crop_new_height, crop_new_width, 3), dtype=np.uint8) * 255
+    if isinstance(input_img, np.ndarray):
 
-    # Crop the original image using numpy slicing
-    cropped_img = input_np_img[crop_ymin:crop_ymax, crop_xmin:crop_xmax]
+        # Create a white background array
+        return_image = np.ones((crop_new_height, crop_new_width, 3), dtype=np.uint8) * 255
 
-    # Paste the cropped image onto the white background
-    return_image[crop_paste_y:crop_paste_y + (crop_ymax - crop_ymin),
-    crop_paste_x:crop_paste_x + (crop_xmax - crop_xmin)] = cropped_img
+        # Crop the original image using numpy slicing
+        cropped_img = input_img[crop_ymin:crop_ymax, crop_xmin:crop_xmax]
+
+        # Paste the cropped image onto the white background
+        return_image[crop_paste_y:crop_paste_y + (crop_ymax - crop_ymin),
+        crop_paste_x:crop_paste_x + (crop_xmax - crop_xmin)] = cropped_img
+    else:
+        # Create a white background array
+        return_image = Image.new('RGB', (crop_new_width, crop_new_height), 'white')
+        # Crop image
+        crop_box = (crop_xmin, crop_ymin, crop_xmax, crop_ymax)
+        cropped_img = input_img.crop(crop_box)
+        return_image.paste(cropped_img, (crop_paste_x, crop_paste_y))
 
     return_list = [crop_paste_x, crop_paste_y, crop_xmin, crop_ymin, crop_xmax, crop_ymax, crop_new_width,
                    crop_new_height]
@@ -287,6 +297,20 @@ def get_res_list_from_layout_res(layout_res, iou_threshold=0.7, overlap_threshol
     return ocr_res_list, filtered_table_res_list, single_page_mfdetrec_res
 
 
+def clean_memory(device='cuda'):
+    if device == 'cuda':
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+            torch.cuda.ipc_collect()
+    elif str(device).startswith("npu"):
+        import torch_npu
+        if torch_npu.npu.is_available():
+            torch_npu.npu.empty_cache()
+    elif str(device).startswith("mps"):
+        torch.mps.empty_cache()
+    gc.collect()
+
+
 def clean_vram(device, vram_threshold=8):
     total_memory = get_vram(device)
     if total_memory and total_memory <= vram_threshold:

mineru/utils/models_download_utils.py

+import os
+from huggingface_hub import snapshot_download as hf_snapshot_download
+from modelscope import snapshot_download as ms_snapshot_download
+
+from mineru.utils.config_reader import get_local_models_dir
+from mineru.utils.enum_class import ModelPath
+
+def auto_download_and_get_model_root_path(relative_path: str, repo_mode='pipeline') -> str:
+    """
+    支持文件或目录的可靠下载。
+    - 如果输入文件: 返回本地文件绝对路径
+    - 如果输入目录: 返回本地缓存下与 relative_path 同结构的相对路径字符串
+    :param repo_mode: 指定仓库模式，'pipeline' 或 'vlm'
+    :param relative_path: 文件或目录相对路径
+    :return: 本地文件绝对路径或相对路径
+    """
+    model_source = os.getenv('MINERU_MODEL_SOURCE', "huggingface")
+
+    if model_source == 'local':
+        local_models_config = get_local_models_dir()
+        root_path = local_models_config.get(repo_mode, None)
+        if not root_path:
+            raise ValueError(f"Local path for repo_mode '{repo_mode}' is not configured.")
+        return root_path
+
+    # 建立仓库模式到路径的映射
+    repo_mapping = {
+        'pipeline': {
+            'huggingface': ModelPath.pipeline_root_hf,
+            'modelscope': ModelPath.pipeline_root_modelscope,
+            'default': ModelPath.pipeline_root_hf
+        },
+        'vlm': {
+            'huggingface': ModelPath.vlm_root_hf,
+            'modelscope': ModelPath.vlm_root_modelscope,
+            'default': ModelPath.vlm_root_hf
+        }
+    }
+
+    if repo_mode not in repo_mapping:
+        raise ValueError(f"Unsupported repo_mode: {repo_mode}, must be 'pipeline' or 'vlm'")
+
+    # 如果没有指定model_source或值不是'modelscope'，则使用默认值
+    repo = repo_mapping[repo_mode].get(model_source, repo_mapping[repo_mode]['default'])
+
+
+    if model_source == "huggingface":
+        snapshot_download = hf_snapshot_download
+    elif model_source == "modelscope":
+        snapshot_download = ms_snapshot_download
+    else:
+        raise ValueError(f"未知的仓库类型: {model_source}")
+
+    cache_dir = None
+
+    if repo_mode == 'pipeline':
+        relative_path = relative_path.strip('/')
+        cache_dir = snapshot_download(repo, allow_patterns=[relative_path, relative_path+"/*"])
+    elif repo_mode == 'vlm':
+        # VLM 模式下，直接下载整个模型目录
+        cache_dir = snapshot_download(repo)
+
+    if not cache_dir:
+        raise FileNotFoundError(f"Failed to download model: {relative_path} from {repo}")
+    return cache_dir
+
+
+if __name__ == '__main__':
+    path1 = "models/README.md"
+    root = auto_download_and_get_model_root_path(path1)
+    print("本地文件绝对路径:", os.path.join(root, path1))
\ No newline at end of file

magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/ocr_utils.py → mineru/utils/ocr_utils.py

 # Copyright (c) Opendatalab. All rights reserved.
 import copy
-
 import cv2
 import numpy as np
-from magic_pdf.pre_proc.ocr_dict_merge import merge_spans_to_line
-from magic_pdf.libs.boxbase import __is_overlaps_y_exceeds_threshold
+
+
+class OcrConfidence:
+    min_confidence = 0.68
+    min_width = 3
+
+
+def merge_spans_to_line(spans, threshold=0.6):
+    if len(spans) == 0:
+        return []
+    else:
+        # 按照y0坐标排序
+        spans.sort(key=lambda span: span['bbox'][1])
+
+        lines = []
+        current_line = [spans[0]]
+        for span in spans[1:]:
+            # 如果当前的span与当前行的最后一个span在y轴上重叠，则添加到当前行
+            if __is_overlaps_y_exceeds_threshold(span['bbox'], current_line[-1]['bbox'], threshold):
+                current_line.append(span)
+            else:
+                # 否则，开始新行
+                lines.append(current_line)
+                current_line = [span]
+
+        # 添加最后一行
+        if current_line:
+            lines.append(current_line)
+
+        return lines
+
+def __is_overlaps_y_exceeds_threshold(bbox1,
+                                      bbox2,
+                                      overlap_ratio_threshold=0.8):
+    """检查两个bbox在y轴上是否有重叠，并且该重叠区域的高度占两个bbox高度更低的那个超过80%"""
+    _, y0_1, _, y1_1 = bbox1
+    _, y0_2, _, y1_2 = bbox2
+
+    overlap = max(0, min(y1_1, y1_2) - max(y0_1, y0_2))
+    height1, height2 = y1_1 - y0_1, y1_2 - y0_2
+    # max_height = max(height1, height2)
+    min_height = min(height1, height2)
+
+    return (overlap / min_height) > overlap_ratio_threshold
 
 
 def img_decode(content: bytes):
@@ -212,10 +253,7 @@ def merge_det_boxes(dt_boxes):
             angle_boxes_list.append(text_box)
             continue
 
-        text_box_dict = {
-            'bbox': text_bbox,
-            'type': 'text',
-        }
+        text_box_dict = {'bbox': text_bbox}
         dt_boxes_dict_list.append(text_box_dict)
 
     # Merge adjacent text regions into lines
@@ -271,7 +309,7 @@ def get_ocr_result_list(ocr_res, useful_list, ocr_enable, new_image, lang):
             p1, p2, p3, p4 = box_ocr_res[0]
             text, score = box_ocr_res[1]
             # logger.info(f"text: {text}, score: {score}")
-            if score < 0.6:  # 过滤低置信度的结果
+            if score < OcrConfidence.min_confidence:  # 过滤低置信度的结果
                 continue
         else:
             p1, p2, p3, p4 = box_ocr_res
@@ -284,6 +322,11 @@ def get_ocr_result_list(ocr_res, useful_list, ocr_enable, new_image, lang):
         # average_angle_degrees = calculate_angle_degrees(box_ocr_res[0])
         # if average_angle_degrees > 0.5:
         poly = [p1, p2, p3, p4]
+
+        if (p3[0] - p1[0]) < OcrConfidence.min_width:
+            # logger.info(f"width too small: {p3[0] - p1[0]}, text: {text}")
+            continue
+
         if calculate_is_angle(poly):
             # logger.info(f"average_angle_degrees: {average_angle_degrees}, text: {text}")
             # 与x轴的夹角超过0.5度，对边界做一下矫正

mineru/utils/pdf_classify.py

+# Copyright (c) Opendatalab. All rights reserved.
+import re
+from io import BytesIO
+import numpy as np
+import pypdfium2 as pdfium
+from loguru import logger
+from pdfminer.high_level import extract_text
+from pdfminer.pdfparser import PDFParser
+from pdfminer.pdfdocument import PDFDocument
+from pdfminer.pdfpage import PDFPage
+from pdfminer.pdfinterp import PDFResourceManager
+from pdfminer.pdfinterp import PDFPageInterpreter
+from pdfminer.layout import LAParams, LTImage, LTFigure
+from pdfminer.converter import PDFPageAggregator
+
+
+def classify(pdf_bytes):
+    """
+    判断PDF文件是可以直接提取文本还是需要OCR
+
+    Args:
+        pdf_bytes: PDF文件的字节数据
+
+    Returns:
+        str: 'txt' 表示可以直接提取文本，'ocr' 表示需要OCR
+    """
+    try:
+        # 从字节数据加载PDF
+        sample_pdf_bytes = extract_pages(pdf_bytes)
+        pdf = pdfium.PdfDocument(sample_pdf_bytes)
+
+        # 获取PDF页数
+        page_count = len(pdf)
+
+        # 如果PDF页数为0，直接返回OCR
+        if page_count == 0:
+            return 'ocr'
+
+        # 检查的页面数（最多检查10页）
+        pages_to_check = min(page_count, 10)
+
+        # 设置阈值：如果每页平均少于50个有效字符，认为需要OCR
+        chars_threshold = 50
+
+        if (get_avg_cleaned_chars_per_page(pdf, pages_to_check) < chars_threshold) or detect_invalid_chars(sample_pdf_bytes):
+            return 'ocr'
+        else:
+
+            if get_high_image_coverage_ratio(sample_pdf_bytes, pages_to_check) >= 0.8:
+                return 'ocr'
+
+            return 'txt'
+    except Exception as e:
+        logger.error(f"判断PDF类型时出错: {e}")
+        # 出错时默认使用OCR
+        return 'ocr'
+
+
+def get_avg_cleaned_chars_per_page(pdf_doc, pages_to_check):
+    # 总字符数
+    total_chars = 0
+    # 清理后的总字符数
+    cleaned_total_chars = 0
+
+    # 检查前几页的文本
+    for i in range(pages_to_check):
+        page = pdf_doc[i]
+        text_page = page.get_textpage()
+        text = text_page.get_text_bounded()
+        total_chars += len(text)
+
+        # 清理提取的文本，移除空白字符
+        cleaned_text = re.sub(r'\s+', '', text)
+        cleaned_total_chars += len(cleaned_text)
+
+    # 计算平均每页字符数
+    avg_cleaned_chars_per_page = cleaned_total_chars / pages_to_check
+
+    # logger.debug(f"PDF分析: 平均每页清理后{avg_cleaned_chars_per_page:.1f}字符")
+
+    pdf_doc.close()  # 关闭PDF文档
+
+    return avg_cleaned_chars_per_page
+
+
+def get_high_image_coverage_ratio(sample_pdf_bytes, pages_to_check):
+    # 创建内存文件对象
+    pdf_stream = BytesIO(sample_pdf_bytes)
+
+    # 创建PDF解析器
+    parser = PDFParser(pdf_stream)
+
+    # 创建PDF文档对象
+    document = PDFDocument(parser)
+
+    # 检查文档是否允许文本提取
+    if not document.is_extractable:
+        # logger.warning("PDF不允许内容提取")
+        return 1.0  # 默认为高覆盖率，因为无法提取内容
+
+    # 创建资源管理器和参数对象
+    rsrcmgr = PDFResourceManager()
+    laparams = LAParams(
+        line_overlap=0.5,
+        char_margin=2.0,
+        line_margin=0.5,
+        word_margin=0.1,
+        boxes_flow=None,
+        detect_vertical=False,
+        all_texts=False,
+    )
+
+    # 创建聚合器
+    device = PDFPageAggregator(rsrcmgr, laparams=laparams)
+
+    # 创建解释器
+    interpreter = PDFPageInterpreter(rsrcmgr, device)
+
+    # 记录高图像覆盖率的页面数量
+    high_image_coverage_pages = 0
+    page_count = 0
+
+    # 遍历页面
+    for page in PDFPage.create_pages(document):
+        # 控制检查的页数
+        if page_count >= pages_to_check:
+            break
+
+        # 处理页面
+        interpreter.process_page(page)
+        layout = device.get_result()
+
+        # 页面尺寸
+        page_width = layout.width
+        page_height = layout.height
+        page_area = page_width * page_height
+
+        # 计算图像覆盖的总面积
+        image_area = 0
+
+        # 遍历页面元素
+        for element in layout:
+            # 检查是否为图像或图形元素
+            if isinstance(element, (LTImage, LTFigure)):
+                # 计算图像边界框面积
+                img_width = element.width
+                img_height = element.height
+                img_area = img_width * img_height
+                image_area += img_area
+
+        # 计算覆盖率
+        coverage_ratio = min(image_area / page_area, 1.0) if page_area > 0 else 0
+        # logger.debug(f"PDF分析: 页面 {page_count + 1} 图像覆盖率: {coverage_ratio:.2f}")
+
+        # 判断是否为高覆盖率
+        if coverage_ratio >= 0.8:  # 使用80%作为高覆盖率的阈值
+            high_image_coverage_pages += 1
+
+        page_count += 1
+
+    # 如果没有处理任何页面，返回0
+    if page_count == 0:
+        return 0.0
+
+    # 计算高图像覆盖率的页面比例
+    high_coverage_ratio = high_image_coverage_pages / page_count
+    # logger.debug(f"PDF分析: 高图像覆盖页面比例: {high_coverage_ratio:.2f}")
+
+    # 关闭资源
+    pdf_stream.close()
+
+    return high_coverage_ratio
+
+
+def extract_pages(src_pdf_bytes: bytes) -> bytes:
+    """
+    从PDF字节数据中随机提取最多10页，返回新的PDF字节数据
+
+    Args:
+        src_pdf_bytes: PDF文件的字节数据
+
+    Returns:
+        bytes: 提取页面后的PDF字节数据
+    """
+
+    # 从字节数据加载PDF
+    pdf = pdfium.PdfDocument(src_pdf_bytes)
+
+    # 获取PDF页数
+    total_page = len(pdf)
+    if total_page == 0:
+        # 如果PDF没有页面，直接返回空文档
+        logger.warning("PDF is empty, return empty document")
+        return b''
+
+    # 选择最多10页
+    select_page_cnt = min(10, total_page)
+
+    # 从总页数中随机选择页面
+    page_indices = np.random.choice(total_page, select_page_cnt, replace=False).tolist()
+
+    # 创建一个新的PDF文档
+    sample_docs = pdfium.PdfDocument.new()
+
+    try:
+        # 将选择的页面导入新文档
+        sample_docs.import_pages(pdf, page_indices)
+
+        # 将新PDF保存到内存缓冲区
+        output_buffer = BytesIO()
+        sample_docs.save(output_buffer)
+
+        # 获取字节数据
+        return output_buffer.getvalue()
+    except Exception as e:
+        logger.exception(e)
+        return b''  # 出错时返回空字节
+
+
+def detect_invalid_chars(sample_pdf_bytes: bytes) -> bool:
+    """"
+    检测PDF中是否包含非法字符
+    """
+    '''pdfminer比较慢,需要先随机抽取10页左右的sample'''
+    # sample_pdf_bytes = extract_pages(src_pdf_bytes)
+    sample_pdf_file_like_object = BytesIO(sample_pdf_bytes)
+    laparams = LAParams(
+        line_overlap=0.5,
+        char_margin=2.0,
+        line_margin=0.5,
+        word_margin=0.1,
+        boxes_flow=None,
+        detect_vertical=False,
+        all_texts=False,
+    )
+    text = extract_text(pdf_file=sample_pdf_file_like_object, laparams=laparams)
+    text = text.replace("\n", "")
+    # logger.info(text)
+    '''乱码文本用pdfminer提取出来的文本特征是(cid:xxx)'''
+    cid_pattern = re.compile(r'\(cid:\d+\)')
+    matches = cid_pattern.findall(text)
+    cid_count = len(matches)
+    cid_len = sum(len(match) for match in matches)
+    text_len = len(text)
+    if text_len == 0:
+        cid_chars_radio = 0
+    else:
+        cid_chars_radio = cid_count/(cid_count + text_len - cid_len)
+    # logger.debug(f"cid_count: {cid_count}, text_len: {text_len}, cid_chars_radio: {cid_chars_radio}")
+    '''当一篇文章存在5%以上的文本是乱码时,认为该文档为乱码文档'''
+    if cid_chars_radio > 0.05:
+        return True  # 乱码文档
+    else:
+        return False   # 正常文档
+
+
+if __name__ == '__main__':
+    with open('/Users/myhloli/pdf/luanma2x10.pdf', 'rb') as f:
+        p_bytes = f.read()
+        logger.info(f"PDF分类结果: {classify(p_bytes)}")
\ No newline at end of file

mineru/utils/pdf_image_tools.py

+# Copyright (c) Opendatalab. All rights reserved.
+from io import BytesIO
+
+import pypdfium2 as pdfium
+from loguru import logger
+from PIL import Image
+
+from mineru.data.data_reader_writer import FileBasedDataWriter
+from mineru.utils.pdf_reader import image_to_b64str, image_to_bytes, page_to_image
+from .hash_utils import str_sha256
+
+
+def pdf_page_to_image(page: pdfium.PdfPage, dpi=200) -> dict:
+    """Convert pdfium.PdfDocument to image, Then convert the image to base64.
+
+    Args:
+        page (_type_): pdfium.PdfPage
+        dpi (int, optional): reset the dpi of dpi. Defaults to 200.
+
+    Returns:
+        dict:  {'img_base64': str, 'img_pil': pil_img, 'scale': float }
+    """
+    pil_img, scale = page_to_image(page, dpi=dpi)
+    img_base64 = image_to_b64str(pil_img)
+
+    image_dict = {
+        "img_base64": img_base64,
+        "img_pil": pil_img,
+        "scale": scale,
+    }
+    return image_dict
+
+
+def load_images_from_pdf(
+    pdf_bytes: bytes,
+    dpi=200,
+    start_page_id=0,
+    end_page_id=None,
+):
+    images_list = []
+    pdf_doc = pdfium.PdfDocument(pdf_bytes)
+    pdf_page_num = len(pdf_doc)
+    end_page_id = end_page_id if end_page_id is not None and end_page_id >= 0 else pdf_page_num - 1
+    if end_page_id > pdf_page_num - 1:
+        logger.warning("end_page_id is out of range, use images length")
+        end_page_id = pdf_page_num - 1
+
+    for index in range(0, pdf_page_num):
+        if start_page_id <= index <= end_page_id:
+            page = pdf_doc[index]
+            image_dict = pdf_page_to_image(page, dpi=dpi)
+            images_list.append(image_dict)
+
+    return images_list, pdf_doc
+
+
+def cut_image(bbox: tuple, page_num: int, page_pil_img, return_path, image_writer: FileBasedDataWriter, scale=2):
+    """从第page_num页的page中，根据bbox进行裁剪出一张jpg图片，返回图片路径 save_path：需要同时支持s3和本地,
+    图片存放在save_path下，文件名是:
+    {page_num}_{bbox[0]}_{bbox[1]}_{bbox[2]}_{bbox[3]}.jpg , bbox内数字取整。"""
+
+    # 拼接文件名
+    filename = f"{page_num}_{int(bbox[0])}_{int(bbox[1])}_{int(bbox[2])}_{int(bbox[3])}"
+
+    # 老版本返回不带bucket的路径
+    img_path = f"{return_path}_{filename}" if return_path is not None else None
+
+    # 新版本生成平铺路径
+    img_hash256_path = f"{str_sha256(img_path)}.jpg"
+    # img_hash256_path = f'{img_path}.jpg'
+
+    crop_img = get_crop_img(bbox, page_pil_img, scale=scale)
+
+    img_bytes = image_to_bytes(crop_img, image_format="JPEG")
+
+    image_writer.write(img_hash256_path, img_bytes)
+    return img_hash256_path
+
+
+def get_crop_img(bbox: tuple, pil_img, scale=2):
+    scale_bbox = (
+        int(bbox[0] * scale),
+        int(bbox[1] * scale),
+        int(bbox[2] * scale),
+        int(bbox[3] * scale),
+    )
+    return pil_img.crop(scale_bbox)
+
+
+def images_bytes_to_pdf_bytes(image_bytes):
+    # 内存缓冲区
+    pdf_buffer = BytesIO()
+
+    # 载入并转换所有图像为 RGB 模式
+    image = Image.open(BytesIO(image_bytes)).convert("RGB")
+
+    # 第一张图保存为 PDF，其余追加
+    image.save(pdf_buffer, format="PDF", save_all=True)
+
+    # 获取 PDF bytes 并重置指针（可选）
+    pdf_bytes = pdf_buffer.getvalue()
+    pdf_buffer.close()
+    return pdf_bytes

mineru/utils/pdf_reader.py

+# Copyright (c) Opendatalab. All rights reserved.
+import base64
+from io import BytesIO
+
+from loguru import logger
+from PIL import Image
+from pypdfium2 import PdfBitmap, PdfDocument, PdfPage
+
+
+def page_to_image(
+    page: PdfPage,
+    dpi: int = 144,  # changed from 200 to 144
+    max_width_or_height: int = 2560,  # changed from 4500 to 2560
+) -> (Image.Image, float):
+    scale = dpi / 72
+
+    long_side_length = max(*page.get_size())
+    if long_side_length > max_width_or_height:
+        scale = max_width_or_height / long_side_length
+
+    bitmap: PdfBitmap = page.render(scale=scale)  # type: ignore
+    try:
+        image = bitmap.to_pil()
+    finally:
+        try:
+            bitmap.close()
+        except Exception:
+            pass
+    return image, scale
+
+
+def image_to_bytes(
+    image: Image.Image,
+    image_format: str = "PNG",  # 也可以用 "JPEG"
+) -> bytes:
+    with BytesIO() as image_buffer:
+        image.save(image_buffer, format=image_format)
+        return image_buffer.getvalue()
+
+
+def image_to_b64str(
+    image: Image.Image,
+    image_format: str = "PNG",  # 也可以用 "JPEG"
+) -> str:
+    image_bytes = image_to_bytes(image, image_format)
+    return base64.b64encode(image_bytes).decode("utf-8")
+
+
+def pdf_to_images(
+    pdf: str | bytes | PdfDocument,
+    dpi: int = 144,
+    max_width_or_height: int = 2560,
+    start_page_id: int = 0,
+    end_page_id: int | None = None,
+) -> list[Image.Image]:
+    doc = pdf if isinstance(pdf, PdfDocument) else PdfDocument(pdf)
+    page_num = len(doc)
+
+    end_page_id = end_page_id if end_page_id is not None and end_page_id >= 0 else page_num - 1
+    if end_page_id > page_num - 1:
+        logger.warning("end_page_id is out of range, use images length")
+        end_page_id = page_num - 1
+
+    images = []
+    try:
+        for i in range(start_page_id, end_page_id + 1):
+            image, _ = page_to_image(doc[i], dpi, max_width_or_height)
+            images.append(image)
+    finally:
+        try:
+            doc.close()
+        except Exception:
+            pass
+    return images
+
+
+def pdf_to_images_bytes(
+    pdf: str | bytes | PdfDocument,
+    dpi: int = 144,
+    max_width_or_height: int = 2560,
+    start_page_id: int = 0,
+    end_page_id: int | None = None,
+    image_format: str = "PNG",
+) -> list[bytes]:
+    images = pdf_to_images(pdf, dpi, max_width_or_height, start_page_id, end_page_id)
+    return [image_to_bytes(image, image_format) for image in images]
+
+
+def pdf_to_images_b64strs(
+    pdf: str | bytes | PdfDocument,
+    dpi: int = 144,
+    max_width_or_height: int = 2560,
+    start_page_id: int = 0,
+    end_page_id: int | None = None,
+    image_format: str = "PNG",
+) -> list[str]:
+    images = pdf_to_images(pdf, dpi, max_width_or_height, start_page_id, end_page_id)
+    return [image_to_b64str(image, image_format) for image in images]

mineru/utils/pdf_text_tool.py

+from typing import List
+import math
+
+import pypdfium2 as pdfium
+from pdftext.pdf.chars import get_chars, deduplicate_chars
+from pdftext.pdf.pages import get_spans, get_lines, assign_scripts, get_blocks
+
+
+def get_page(
+    page: pdfium.PdfPage,
+    quote_loosebox: bool =True,
+    superscript_height_threshold: float = 0.7,
+    line_distance_threshold: float = 0.1,
+) -> dict:
+
+        textpage = page.get_textpage()
+        page_bbox: List[float] = page.get_bbox()
+        page_width = math.ceil(abs(page_bbox[2] - page_bbox[0]))
+        page_height = math.ceil(abs(page_bbox[1] - page_bbox[3]))
+
+        page_rotation = 0
+        try:
+            page_rotation = page.get_rotation()
+        except:
+            pass
+
+        chars = deduplicate_chars(get_chars(textpage, page_bbox, page_rotation, quote_loosebox))
+        spans = get_spans(chars, superscript_height_threshold=superscript_height_threshold, line_distance_threshold=line_distance_threshold)
+        lines = get_lines(spans)
+        assign_scripts(lines, height_threshold=superscript_height_threshold, line_distance_threshold=line_distance_threshold)
+        blocks = get_blocks(lines)
+
+        page = {
+            "bbox": page_bbox,
+            "width": page_width,
+            "height": page_height,
+            "rotation": page_rotation,
+            "blocks": blocks
+        }
+        return page
\ No newline at end of file

mineru/utils/run_async.py

+import asyncio
+import threading
+from queue import Queue
+from typing import Any, AsyncIterable, Coroutine, Iterable, TypeVar
+
+T = TypeVar("T")
+
+
+def run_async(coroutine: Coroutine[Any, Any, T]) -> T:
+    if not asyncio.iscoroutine(coroutine):
+        raise ValueError("a coroutine was expected, got {!r}".format(coroutine))
+
+    try:
+        loop = asyncio.get_running_loop()
+    except RuntimeError:
+        loop = None
+
+    if loop is not None:
+        return loop.run_until_complete(coroutine)
+    else:
+        return asyncio.run(coroutine)
+
+
+def iter_async(iterable: AsyncIterable[T]) -> Iterable[T]:
+    if not isinstance(iterable, AsyncIterable):
+        raise ValueError("an async iterable was expected, got {!r}".format(iterable))
+
+    queue = Queue()
+
+    async def async_helper():
+        try:
+            async for chunk in iterable:
+                queue.put(chunk)
+            queue.put(None)
+        except Exception as e:
+            queue.put(e)
+
+    def helper():
+        run_async(async_helper())
+
+    thread = threading.Thread(target=helper, daemon=True)
+    thread.start()
+
+    while True:
+        chunk = queue.get()
+        if chunk is None:
+            break
+        if isinstance(chunk, Exception):
+            raise chunk
+        yield chunk
+
+    thread.join()

magic_pdf/pre_proc/ocr_dict_merge.py → mineru/utils/span_block_fix.py

-from magic_pdf.config.ocr_content_type import BlockType, ContentType
-from magic_pdf.libs.boxbase import __is_overlaps_y_exceeds_threshold, calculate_overlap_area_in_bbox1_area_ratio
+# Copyright (c) Opendatalab. All rights reserved.
+from mineru.utils.boxbase import calculate_overlap_area_in_bbox1_area_ratio
+from mineru.utils.enum_class import BlockType, ContentType
+from mineru.utils.ocr_utils import __is_overlaps_y_exceeds_threshold
 
 
-# 将每一个line中的span从左到右排序
-def line_sort_spans_by_left_to_right(lines):
-    line_objects = []
-    for line in lines:
-        #  按照x0坐标排序
-        line.sort(key=lambda span: span['bbox'][0])
-        line_bbox = [
-            min(span['bbox'][0] for span in line),  # x0
-            min(span['bbox'][1] for span in line),  # y0
-            max(span['bbox'][2] for span in line),  # x1
-            max(span['bbox'][3] for span in line),  # y1
+def fill_spans_in_blocks(blocks, spans, radio):
+    """将allspans中的span按位置关系，放入blocks中."""
+    block_with_spans = []
+    for block in blocks:
+        block_type = block[7]
+        block_bbox = block[0:4]
+        block_dict = {
+            'type': block_type,
+            'bbox': block_bbox,
+        }
+        if block_type in [
+            BlockType.IMAGE_BODY, BlockType.IMAGE_CAPTION, BlockType.IMAGE_FOOTNOTE,
+            BlockType.TABLE_BODY, BlockType.TABLE_CAPTION, BlockType.TABLE_FOOTNOTE
+        ]:
+            block_dict['group_id'] = block[-1]
+        block_spans = []
+        for span in spans:
+            span_bbox = span['bbox']
+            if calculate_overlap_area_in_bbox1_area_ratio(span_bbox, block_bbox) > radio and span_block_type_compatible(
+                    span['type'], block_type):
+                block_spans.append(span)
+
+        block_dict['spans'] = block_spans
+        block_with_spans.append(block_dict)
+
+        # 从spans删除已经放入block_spans中的span
+        if len(block_spans) > 0:
+            for span in block_spans:
+                spans.remove(span)
+
+    return block_with_spans, spans
+
+
+def span_block_type_compatible(span_type, block_type):
+    if span_type in [ContentType.TEXT, ContentType.INLINE_EQUATION]:
+        return block_type in [
+            BlockType.TEXT,
+            BlockType.TITLE,
+            BlockType.IMAGE_CAPTION,
+            BlockType.IMAGE_FOOTNOTE,
+            BlockType.TABLE_CAPTION,
+            BlockType.TABLE_FOOTNOTE,
+            BlockType.DISCARDED
         ]
-        line_objects.append({
-            'bbox': line_bbox,
-            'spans': line,
-        })
-    return line_objects
+    elif span_type == ContentType.INTERLINE_EQUATION:
+        return block_type in [BlockType.INTERLINE_EQUATION, BlockType.TEXT]
+    elif span_type == ContentType.IMAGE:
+        return block_type in [BlockType.IMAGE_BODY]
+    elif span_type == ContentType.TABLE:
+        return block_type in [BlockType.TABLE_BODY]
+    else:
+        return False
+
+
+def fix_discarded_block(discarded_block_with_spans):
+    fix_discarded_blocks = []
+    for block in discarded_block_with_spans:
+        block = fix_text_block(block)
+        fix_discarded_blocks.append(block)
+    return fix_discarded_blocks
+
+
+def fix_text_block(block):
+    # 文本block中的公式span都应该转换成行内type
+    for span in block['spans']:
+        if span['type'] == ContentType.INTERLINE_EQUATION:
+            span['type'] = ContentType.INLINE_EQUATION
+    block_lines = merge_spans_to_line(block['spans'])
+    sort_block_lines = line_sort_spans_by_left_to_right(block_lines)
+    block['lines'] = sort_block_lines
+    del block['spans']
+    return block
 
 
 def merge_spans_to_line(spans, threshold=0.6):
@@ -34,11 +91,11 @@ def merge_spans_to_line(spans, threshold=0.6):
             # 如果当前的span类型为"interline_equation" 或者 当前行中已经有"interline_equation"
             # image和table类型，同上
             if span['type'] in [
-                    ContentType.InterlineEquation, ContentType.Image,
-                    ContentType.Table
+                    ContentType.INTERLINE_EQUATION, ContentType.IMAGE,
+                    ContentType.TABLE
             ] or any(s['type'] in [
-                    ContentType.InterlineEquation, ContentType.Image,
-                    ContentType.Table
+                    ContentType.INTERLINE_EQUATION, ContentType.IMAGE,
+                    ContentType.TABLE
             ] for s in current_line):
                 # 则开始新行
                 lines.append(current_line)
@@ -60,70 +117,36 @@ def merge_spans_to_line(spans, threshold=0.6):
         return lines
 
 
-def span_block_type_compatible(span_type, block_type):
-    if span_type in [ContentType.Text, ContentType.InlineEquation]:
-        return block_type in [
-            BlockType.Text,
-            BlockType.Title,
-            BlockType.ImageCaption,
-            BlockType.ImageFootnote,
-            BlockType.TableCaption,
-            BlockType.TableFootnote,
-            BlockType.Discarded
+# 将每一个line中的span从左到右排序
+def line_sort_spans_by_left_to_right(lines):
+    line_objects = []
+    for line in lines:
+        #  按照x0坐标排序
+        line.sort(key=lambda span: span['bbox'][0])
+        line_bbox = [
+            min(span['bbox'][0] for span in line),  # x0
+            min(span['bbox'][1] for span in line),  # y0
+            max(span['bbox'][2] for span in line),  # x1
+            max(span['bbox'][3] for span in line),  # y1
         ]
-    elif span_type == ContentType.InterlineEquation:
-        return block_type in [BlockType.InterlineEquation, BlockType.Text]
-    elif span_type == ContentType.Image:
-        return block_type in [BlockType.ImageBody]
-    elif span_type == ContentType.Table:
-        return block_type in [BlockType.TableBody]
-    else:
-        return False
-
-
-def fill_spans_in_blocks(blocks, spans, radio):
-    """将allspans中的span按位置关系，放入blocks中."""
-    block_with_spans = []
-    for block in blocks:
-        block_type = block[7]
-        block_bbox = block[0:4]
-        block_dict = {
-            'type': block_type,
-            'bbox': block_bbox,
-        }
-        if block_type in [
-            BlockType.ImageBody, BlockType.ImageCaption, BlockType.ImageFootnote,
-            BlockType.TableBody, BlockType.TableCaption, BlockType.TableFootnote
-        ]:
-            block_dict['group_id'] = block[-1]
-        block_spans = []
-        for span in spans:
-            span_bbox = span['bbox']
-            if calculate_overlap_area_in_bbox1_area_ratio(span_bbox, block_bbox) > radio and span_block_type_compatible(span['type'], block_type):
-                block_spans.append(span)
-
-        block_dict['spans'] = block_spans
-        block_with_spans.append(block_dict)
-
-        # 从spans删除已经放入block_spans中的span
-        if len(block_spans) > 0:
-            for span in block_spans:
-                spans.remove(span)
-
-    return block_with_spans, spans
+        line_objects.append({
+            'bbox': line_bbox,
+            'spans': line,
+        })
+    return line_objects
 
 
-def fix_block_spans_v2(block_with_spans):
+def fix_block_spans(block_with_spans):
     fix_blocks = []
     for block in block_with_spans:
         block_type = block['type']
 
-        if block_type in [BlockType.Text, BlockType.Title,
-                          BlockType.ImageCaption, BlockType.ImageFootnote,
-                          BlockType.TableCaption, BlockType.TableFootnote
+        if block_type in [BlockType.TEXT, BlockType.TITLE,
+                          BlockType.IMAGE_CAPTION, BlockType.IMAGE_CAPTION,
+                          BlockType.TABLE_CAPTION, BlockType.TABLE_FOOTNOTE
                           ]:
             block = fix_text_block(block)
-        elif block_type in [BlockType.InterlineEquation, BlockType.ImageBody, BlockType.TableBody]:
+        elif block_type in [BlockType.INTERLINE_EQUATION, BlockType.IMAGE_BODY, BlockType.TABLE_BODY]:
             block = fix_interline_block(block)
         else:
             continue
@@ -131,29 +154,9 @@ def fix_block_spans_v2(block_with_spans):
     return fix_blocks
 
 
-def fix_discarded_block(discarded_block_with_spans):
-    fix_discarded_blocks = []
-    for block in discarded_block_with_spans:
-        block = fix_text_block(block)
-        fix_discarded_blocks.append(block)
-    return fix_discarded_blocks
-
-
-def fix_text_block(block):
-    # 文本block中的公式span都应该转换成行内type
-    for span in block['spans']:
-        if span['type'] == ContentType.InterlineEquation:
-            span['type'] = ContentType.InlineEquation
-    block_lines = merge_spans_to_line(block['spans'])
-    sort_block_lines = line_sort_spans_by_left_to_right(block_lines)
-    block['lines'] = sort_block_lines
-    del block['spans']
-    return block
-
-
 def fix_interline_block(block):
     block_lines = merge_spans_to_line(block['spans'])
     sort_block_lines = line_sort_spans_by_left_to_right(block_lines)
     block['lines'] = sort_block_lines
     del block['spans']
-    return block
+    return block
\ No newline at end of file

mineru/utils/span_pre_proc.py

+# Copyright (c) Opendatalab. All rights reserved.
+import re
+import statistics
+
+import cv2
+import numpy as np
+from loguru import logger
+
+from mineru.utils.boxbase import calculate_overlap_area_in_bbox1_area_ratio, calculate_iou, \
+    get_minbox_if_overlap_by_ratio
+from mineru.utils.enum_class import BlockType, ContentType
+from mineru.utils.pdf_image_tools import get_crop_img
+from mineru.utils.pdf_text_tool import get_page
+
+
+def remove_outside_spans(spans, all_bboxes, all_discarded_blocks):
+    def get_block_bboxes(blocks, block_type_list):
+        return [block[0:4] for block in blocks if block[7] in block_type_list]
+
+    image_bboxes = get_block_bboxes(all_bboxes, [BlockType.IMAGE_BODY])
+    table_bboxes = get_block_bboxes(all_bboxes, [BlockType.TABLE_BODY])
+    other_block_type = []
+    for block_type in BlockType.__dict__.values():
+        if not isinstance(block_type, str):
+            continue
+        if block_type not in [BlockType.IMAGE_BODY, BlockType.TABLE_BODY]:
+            other_block_type.append(block_type)
+    other_block_bboxes = get_block_bboxes(all_bboxes, other_block_type)
+    discarded_block_bboxes = get_block_bboxes(all_discarded_blocks, [BlockType.DISCARDED])
+
+    new_spans = []
+
+    for span in spans:
+        span_bbox = span['bbox']
+        span_type = span['type']
+
+        if any(calculate_overlap_area_in_bbox1_area_ratio(span_bbox, block_bbox) > 0.4 for block_bbox in
+               discarded_block_bboxes):
+            new_spans.append(span)
+            continue
+
+        if span_type == ContentType.IMAGE:
+            if any(calculate_overlap_area_in_bbox1_area_ratio(span_bbox, block_bbox) > 0.5 for block_bbox in
+                   image_bboxes):
+                new_spans.append(span)
+        elif span_type == ContentType.TABLE:
+            if any(calculate_overlap_area_in_bbox1_area_ratio(span_bbox, block_bbox) > 0.5 for block_bbox in
+                   table_bboxes):
+                new_spans.append(span)
+        else:
+            if any(calculate_overlap_area_in_bbox1_area_ratio(span_bbox, block_bbox) > 0.5 for block_bbox in
+                   other_block_bboxes):
+                new_spans.append(span)
+
+    return new_spans
+
+
+def remove_overlaps_low_confidence_spans(spans):
+    dropped_spans = []
+    #  删除重叠spans中置信度低的的那些
+    for span1 in spans:
+        for span2 in spans:
+            if span1 != span2:
+                # span1 或 span2 任何一个都不应该在 dropped_spans 中
+                if span1 in dropped_spans or span2 in dropped_spans:
+                    continue
+                else:
+                    if calculate_iou(span1['bbox'], span2['bbox']) > 0.9:
+                        if span1['score'] < span2['score']:
+                            span_need_remove = span1
+                        else:
+                            span_need_remove = span2
+                        if (
+                            span_need_remove is not None
+                            and span_need_remove not in dropped_spans
+                        ):
+                            dropped_spans.append(span_need_remove)
+
+    if len(dropped_spans) > 0:
+        for span_need_remove in dropped_spans:
+            spans.remove(span_need_remove)
+
+    return spans, dropped_spans
+
+
+def remove_overlaps_min_spans(spans):
+    dropped_spans = []
+    #  删除重叠spans中较小的那些
+    for span1 in spans:
+        for span2 in spans:
+            if span1 != span2:
+                # span1 或 span2 任何一个都不应该在 dropped_spans 中
+                if span1 in dropped_spans or span2 in dropped_spans:
+                    continue
+                else:
+                    overlap_box = get_minbox_if_overlap_by_ratio(span1['bbox'], span2['bbox'], 0.65)
+                    if overlap_box is not None:
+                        span_need_remove = next((span for span in spans if span['bbox'] == overlap_box), None)
+                        if span_need_remove is not None and span_need_remove not in dropped_spans:
+                            dropped_spans.append(span_need_remove)
+    if len(dropped_spans) > 0:
+        for span_need_remove in dropped_spans:
+            spans.remove(span_need_remove)
+
+    return spans, dropped_spans
+
+
+def __replace_ligatures(text: str):
+    ligatures = {
+        'fi': 'fi', 'fl': 'fl', 'ff': 'ff', 'ffi': 'ffi', 'ffl': 'ffl', 'ſt': 'ft', 'st': 'st'
+    }
+    return re.sub('|'.join(map(re.escape, ligatures.keys())), lambda m: ligatures[m.group()], text)
+
+def __replace_unicode(text: str):
+    ligatures = {
+        '\r\n': '', '\u0002': '-',
+    }
+    return re.sub('|'.join(map(re.escape, ligatures.keys())), lambda m: ligatures[m.group()], text)
+
+
+"""pdf_text dict方案 char级别"""
+def txt_spans_extract(pdf_page, spans, pil_img, scale, all_bboxes, all_discarded_blocks):
+
+    page_dict = get_page(pdf_page)
+
+    page_all_chars = []
+    page_all_lines = []
+    for block in page_dict['blocks']:
+        for line in block['lines']:
+            if 0 < abs(line['rotation']) < 90:
+                # 旋转角度在0-90度之间的行，直接跳过
+                continue
+            page_all_lines.append(line)
+            for span in line['spans']:
+                for char in span['chars']:
+                    page_all_chars.append(char)
+
+    # 计算所有sapn的高度的中位数
+    span_height_list = []
+    for span in spans:
+        if span['type'] in [ContentType.TEXT]:
+            span_height = span['bbox'][3] - span['bbox'][1]
+            span['height'] = span_height
+            span['width'] = span['bbox'][2] - span['bbox'][0]
+            span_height_list.append(span_height)
+    if len(span_height_list) == 0:
+        return spans
+    else:
+        median_span_height = statistics.median(span_height_list)
+
+    useful_spans = []
+    unuseful_spans = []
+    # 纵向span的两个特征：1. 高度超过多个line 2. 高宽比超过某个值
+    vertical_spans = []
+    for span in spans:
+        if span['type'] in [ContentType.TEXT]:
+            for block in all_bboxes + all_discarded_blocks:
+                if block[7] in [BlockType.IMAGE_BODY, BlockType.TABLE_BODY, BlockType.INTERLINE_EQUATION]:
+                    continue
+                if calculate_overlap_area_in_bbox1_area_ratio(span['bbox'], block[0:4]) > 0.5:
+                    if span['height'] > median_span_height * 3 and span['height'] > span['width'] * 3:
+                        vertical_spans.append(span)
+                    elif block in all_bboxes:
+                        useful_spans.append(span)
+                    else:
+                        unuseful_spans.append(span)
+                    break
+
+    """垂直的span框直接用line进行填充"""
+    if len(vertical_spans) > 0:
+        for pdfium_line in page_all_lines:
+            for span in vertical_spans:
+                if calculate_overlap_area_in_bbox1_area_ratio(pdfium_line['bbox'].bbox, span['bbox']) > 0.5:
+                    for pdfium_span in pdfium_line['spans']:
+                        span['content'] += pdfium_span['text']
+                    break
+
+        for span in vertical_spans:
+            if len(span['content']) == 0:
+                spans.remove(span)
+
+    """水平的span框先用char填充，再用ocr填充空的span框"""
+    new_spans = []
+
+    for span in useful_spans + unuseful_spans:
+        if span['type'] in [ContentType.TEXT]:
+            span['chars'] = []
+            new_spans.append(span)
+
+    need_ocr_spans = fill_char_in_spans(new_spans, page_all_chars)
+
+    """对未填充的span进行ocr"""
+    if len(need_ocr_spans) > 0:
+
+        for span in need_ocr_spans:
+            # 对span的bbox截图再ocr
+            span_pil_img = get_crop_img(span['bbox'], pil_img, scale)
+            span_img = cv2.cvtColor(np.array(span_pil_img), cv2.COLOR_RGB2BGR)
+            # 计算span的对比度，低于0.20的span不进行ocr
+            if calculate_contrast(span_img, img_mode='bgr') <= 0.17:
+                spans.remove(span)
+                continue
+
+            span['content'] = ''
+            span['score'] = 1.0
+            span['np_img'] = span_img
+
+    return spans
+
+
+def fill_char_in_spans(spans, all_chars):
+
+    # 简单从上到下排一下序
+    spans = sorted(spans, key=lambda x: x['bbox'][1])
+
+    for char in all_chars:
+
+        for span in spans:
+            if calculate_char_in_span(char['bbox'], span['bbox'], char['char']):
+                span['chars'].append(char)
+                break
+
+    need_ocr_spans = []
+    for span in spans:
+        chars_to_content(span)
+        # 有的span中虽然没有字但有一两个空的占位符，用宽高和content长度过滤
+        if len(span['content']) * span['height'] < span['width'] * 0.5:
+            # logger.info(f"maybe empty span: {len(span['content'])}, {span['height']}, {span['width']}")
+            need_ocr_spans.append(span)
+        del span['height'], span['width']
+    return need_ocr_spans
+
+
+LINE_STOP_FLAG = ('.', '!', '?', '。', '！', '？', ')', '）', '"', '”', ':', '：', ';', '；', ']', '】', '}', '}', '>', '》', '、', ',', '，', '-', '—', '–',)
+LINE_START_FLAG = ('(', '（', '"', '“', '【', '{', '《', '<', '「', '『', '【', '[',)
+
+Span_Height_Radio = 0.33  # 字符的中轴和span的中轴高度差不能超过1/3span高度
+def calculate_char_in_span(char_bbox, span_bbox, char, span_height_radio=Span_Height_Radio):
+    char_center_x = (char_bbox[0] + char_bbox[2]) / 2
+    char_center_y = (char_bbox[1] + char_bbox[3]) / 2
+    span_center_y = (span_bbox[1] + span_bbox[3]) / 2
+    span_height = span_bbox[3] - span_bbox[1]
+
+    if (
+        span_bbox[0] < char_center_x < span_bbox[2]
+        and span_bbox[1] < char_center_y < span_bbox[3]
+        and abs(char_center_y - span_center_y) < span_height * span_height_radio  # 字符的中轴和span的中轴高度差不能超过Span_Height_Radio
+    ):
+        return True
+    else:
+        # 如果char是LINE_STOP_FLAG，就不用中心点判定，换一种方案（左边界在span区域内，高度判定和之前逻辑一致）
+        # 主要是给结尾符号一个进入span的机会，这个char还应该离span右边界较近
+        if char in LINE_STOP_FLAG:
+            if (
+                (span_bbox[2] - span_height) < char_bbox[0] < span_bbox[2]
+                and char_center_x > span_bbox[0]
+                and span_bbox[1] < char_center_y < span_bbox[3]
+                and abs(char_center_y - span_center_y) < span_height * span_height_radio
+            ):
+                return True
+        elif char in LINE_START_FLAG:
+            if (
+                span_bbox[0] < char_bbox[2] < (span_bbox[0] + span_height)
+                and char_center_x < span_bbox[2]
+                and span_bbox[1] < char_center_y < span_bbox[3]
+                and abs(char_center_y - span_center_y) < span_height * span_height_radio
+            ):
+                return True
+        else:
+            return False
+
+
+def chars_to_content(span):
+    # 检查span中的char是否为空
+    if len(span['chars']) == 0:
+        pass
+    else:
+        # 给chars按char_idx排序
+        span['chars'] = sorted(span['chars'], key=lambda x: x['char_idx'])
+
+        # Calculate the width of each character
+        char_widths = [char['bbox'][2] - char['bbox'][0] for char in span['chars']]
+        # Calculate the median width
+        median_width = statistics.median(char_widths)
+
+        content = ''
+        for char in span['chars']:
+
+            # 如果下一个char的x0和上一个char的x1距离超过0.25个字符宽度，则需要在中间插入一个空格
+            char1 = char
+            char2 = span['chars'][span['chars'].index(char) + 1] if span['chars'].index(char) + 1 < len(span['chars']) else None
+            if char2 and char2['bbox'][0] - char1['bbox'][2] > median_width * 0.25 and char['char'] != ' ' and char2['char'] != ' ':
+                content += f"{char['char']} "
+            else:
+                content += char['char']
+
+        content = __replace_unicode(content)
+        content = __replace_ligatures(content)
+        content = __replace_ligatures(content)
+        span['content'] = content.strip()
+
+    del span['chars']
+
+
+def calculate_contrast(img, img_mode) -> float:
+    """
+    计算给定图像的对比度。
+    :param img: 图像，类型为numpy.ndarray
+    :Param img_mode = 图像的色彩通道，'rgb' 或 'bgr'
+    :return: 图像的对比度值
+    """
+    if img_mode == 'rgb':
+        # 将RGB图像转换为灰度图
+        gray_img = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
+    elif img_mode == 'bgr':
+        # 将BGR图像转换为灰度图
+        gray_img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
+    else:
+        raise ValueError("Invalid image mode. Please provide 'rgb' or 'bgr'.")
+
+    # 计算均值和标准差
+    mean_value = np.mean(gray_img)
+    std_dev = np.std(gray_img)
+    # 对比度定义为标准差除以平均值（加上小常数避免除零错误）
+    contrast = std_dev / (mean_value + 1e-6)
+    # logger.debug(f"contrast: {contrast}")
+    return round(contrast, 2)
\ No newline at end of file

mineru/version.py

+__version__ = "2.0.0"
\ No newline at end of file

next_docs/en/.readthedocs.yaml

-version: 2
-
-build:
-  os: ubuntu-22.04
-  tools:
-    python: "3.10"
-
-formats:
-  - epub
-
-python:
-  install:
-    - requirements: next_docs/requirements.txt
-
-sphinx:
-  configuration: next_docs/en/conf.py

next_docs/en/Makefile

-# Minimal makefile for Sphinx documentation
-#
-
-# You can set these variables from the command line, and also
-# from the environment for the first two.
-SPHINXOPTS    ?=
-SPHINXBUILD   ?= sphinx-build
-SOURCEDIR     = .
-BUILDDIR      = _build
-
-# Put it first so that "make" without argument is like "make help".
-help:
-	@$(SPHINXBUILD) -M help "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)
-
-.PHONY: help Makefile
-
-# Catch-all target: route all unknown targets to Sphinx using the new
-# "make mode" option.  $(O) is meant as a shortcut for $(SPHINXOPTS).
-%: Makefile
-	@$(SPHINXBUILD) -M $@ "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)

next_docs/en/_static/image/ReadTheDocs.svg

-<?xml version="1.0" encoding="UTF-8" standalone="no" ?>
-<!DOCTYPE svg PUBLIC "-//W3C//DTD SVG 1.1//EN" "http://www.w3.org/Graphics/SVG/1.1/DTD/svg11.dtd">
-<svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" version="1.1" width="224" height="72" viewBox="-29 -3.67 224 72" xml:space="preserve">
-<desc>Created with Fabric.js 5.2.4</desc>
-<defs>
-</defs>
-<rect x="0" y="0" width="100%" height="100%" fill="transparent"></rect>
-<g transform="matrix(1 0 0 1 112 36)" id="7a867f58-a908-4f30-a839-fb725512b521"  >
-<rect style="stroke: none; stroke-width: 1; stroke-dasharray: none; stroke-linecap: butt; stroke-dashoffset: 0; stroke-linejoin: miter; stroke-miterlimit: 4; fill: rgb(255,255,255); fill-rule: nonzero; opacity: 1; visibility: hidden;" vector-effect="non-scaling-stroke"  x="-112" y="-36" rx="0" ry="0" width="224" height="72" />
-</g>
-<g transform="matrix(Infinity NaN NaN Infinity 0 0)" id="29611287-bf1c-4faf-8eb1-df32f6424829"  >
-</g>
-<g transform="matrix(0.07 0 0 0.07 382.02 122.8)" id="60cdd44f-027a-437a-92c4-c8d44c60ef9e"  >
-<path style="stroke: rgb(0,0,0); stroke-width: 0; stroke-dasharray: none; stroke-linecap: butt; stroke-dashoffset: 0; stroke-linejoin: miter; stroke-miterlimit: 4; fill: rgb(50,50,42); fill-rule: nonzero; opacity: 1;" vector-effect="non-scaling-stroke"  transform=" translate(-64, -64)" d="M 57.62 61.68 C 55.919999999999995 61.92 54.75 63.46 55 65.11 C 55.1668510745875 66.32380621250819 56.039448735907676 67.32218371690155 57.22 67.65 C 57.22 67.65 64.69 70.11 77.4 71.16000000000001 C 87.61000000000001 72.01 99.2 70.43 99.2 70.43 C 100.9 70.39 102.23 68.98 102.19 67.28 C 102.17037752125772 66.4652516996782 101.82707564255573 65.69186585376654 101.23597809465886 65.13079230830253 C 100.644880546762 64.56971876283853 99.85466451370849 64.26716220997277 99.03999999999999 64.29 C 98.83999999999999 64.29 98.63999999999999 64.33000000000001 98.42999999999999 64.37 C 98.42999999999999 64.37 87.08999999999999 65.78 77.88 65.02000000000001 C 65.72999999999999 64.05000000000001 59.11 61.83000000000001 59.11 61.83000000000001 C 58.63 61.670000000000016 58.1 61.59000000000001 57.62 61.670000000000016 Z M 57.62 46.46 C 55.919999999999995 46.7 54.75 48.24 55 49.89 C 55.1668510745875 51.10380621250818 56.039448735907676 52.10218371690154 57.22 52.43 C 57.22 52.43 64.69 54.89 77.4 55.94 C 87.61000000000001 56.79 99.2 55.21 99.2 55.21 C 100.9 55.17 102.23 53.76 102.19 52.06 C 102.17037752125772 51.245251699678214 101.82707564255573 50.47186585376654 101.23597809465886 49.91079230830253 C 100.644880546762 49.34971876283853 99.85466451370849 49.047162209972754 99.03999999999999 49.07 C 98.83999999999999 49.07 98.63999999999999 49.11 98.42999999999999 49.15 C 98.42999999999999 49.15 87.08999999999999 50.559999999999995 77.88 49.8 C 65.72999999999999 48.83 59.11 46.61 59.11 46.61 C 58.63 46.45 58.1 46.37 57.62 46.45 Z M 57.62 31.240000000000002 C 55.919999999999995 31.48 54.75 33.02 55 34.67 C 55.1668510745875 35.88380621250818 56.039448735907676 36.882183716901544 57.22 37.21 C 57.22 37.21 64.69 39.67 77.4 40.72 C 87.61000000000001 41.57 99.2 39.99 99.2 39.99 C 100.9 39.95 102.23 38.54 102.19 36.84 C 102.17037752125772 36.025251699678215 101.82707564255573 35.25186585376654 101.23597809465886 34.690792308302534 C 100.644880546762 34.12971876283853 99.85466451370849 33.827162209972755 99.03999999999999 33.85 C 98.83999999999999 33.85 98.63999999999999 33.89 98.42999999999999 33.93 C 98.42999999999999 33.93 87.08999999999999 35.339999999999996 77.88 34.58 C 65.72999999999999 33.61 59.11 31.389999999999997 59.11 31.389999999999997 C 58.63 31.229999999999997 58.1 31.189999999999998 57.62 31.229999999999997 Z M 57.62 16.060000000000002 C 55.919999999999995 16.3 54.75 17.840000000000003 55 19.490000000000002 C 55.1668510745875 20.703806212508187 56.039448735907676 21.702183716901544 57.22 22.03 C 57.22 22.03 64.69 24.490000000000002 77.4 25.54 C 87.61000000000001 26.39 99.2 24.81 99.2 24.81 C 100.9 24.77 102.23 23.36 102.19 21.66 C 102.17037752125772 20.84525169967821 101.82707564255573 20.07186585376654 101.23597809465886 19.510792308302534 C 100.644880546762 18.949718762838526 99.8546645137085 18.64716220997276 99.03999999999999 18.67 C 98.83999999999999 18.67 98.63999999999999 18.71 98.42999999999999 18.75 C 98.42999999999999 18.75 87.08999999999999 20.16 77.88 19.4 C 65.72999999999999 18.43 59.11 16.209999999999997 59.11 16.209999999999997 C 58.637850878541954 16.01924514007714 58.12188500879498 15.963839409097599 57.62 16.049999999999997 Z M 36.31 0 C 20.32 0.12 14.39 5.05 14.39 5.05 L 14.39 124.42 C 14.39 124.42 20.2 119.41 38.93 120.18 C 57.66 120.95000000000002 61.5 127.53 84.50999999999999 127.97000000000001 C 107.52 128.41000000000003 113.28999999999999 124.42000000000002 113.28999999999999 124.42000000000002 L 113.60999999999999 2.750000000000014 C 113.60999999999999 2.750000000000014 103.28 5.7 83.09 5.86 C 62.95 6.01 58.11 0.73 39.62 0.12 C 38.49 0.04 37.4 0 36.31 0 Z M 49.67 7.79 C 49.67 7.79 59.36 10.98 77.24000000000001 11.870000000000001 C 92.38000000000001 12.64 107.52000000000001 10.38 107.52000000000001 10.38 L 107.52000000000001 118.53 C 107.52000000000001 118.53 99.85000000000001 122.57000000000001 80.68 121.19 C 65.82000000000001 120.14 49.480000000000004 114.49 49.480000000000004 114.49 L 49.68000000000001 7.799999999999997 Z M 40.35 10.620000000000001 C 42.050000000000004 10.620000000000001 43.46 11.990000000000002 43.46 13.73 C 43.46 15.469999999999999 42.09 16.84 40.35 16.84 C 40.35 16.84 35.34 16.88 32.28 17.16 C 27.150000000000002 17.68 23.64 19.54 23.64 19.54 C 22.150000000000002 20.349999999999998 20.25 19.74 19.48 18.25 C 18.67 16.76 19.28 14.86 20.77 14.09 C 22.259999999999998 13.32 25.33 11.67 31.67 11.06 C 35.34 10.66 40.35 10.620000000000001 40.35 10.620000000000001 Z M 37.36 25.880000000000003 C 39.06 25.840000000000003 40.35 25.880000000000003 40.35 25.880000000000003 C 42.050000000000004 26.080000000000002 43.260000000000005 27.62 43.050000000000004 29.310000000000002 C 42.88374644848126 30.726609090871516 41.76660909087151 31.843746448481262 40.35 32.010000000000005 C 40.35 32.010000000000005 35.34 32.050000000000004 32.28 32.330000000000005 C 27.150000000000002 32.85000000000001 23.64 34.71000000000001 23.64 34.71000000000001 C 22.150000000000002 35.52000000000001 20.25 34.91000000000001 19.48 33.42000000000001 C 18.67 31.93000000000001 19.28 30.03000000000001 20.77 29.26000000000001 C 20.77 29.26000000000001 25.33 26.84000000000001 31.67 26.230000000000008 C 33.53 25.99000000000001 35.67 25.910000000000007 37.36 25.870000000000008 Z M 40.35 41.06 C 42.050000000000004 41.06 43.46 42.43 43.46 44.17 C 43.46 45.910000000000004 42.09 47.28 40.35 47.28 C 40.35 47.28 35.34 47.24 32.28 47.56 C 27.150000000000002 48.080000000000005 23.64 49.940000000000005 23.64 49.940000000000005 C 22.150000000000002 50.75000000000001 20.25 50.14000000000001 19.48 48.650000000000006 C 18.67 47.160000000000004 19.28 45.260000000000005 20.77 44.49000000000001 C 20.77 44.49000000000001 25.33 42.07000000000001 31.67 41.46000000000001 C 35.34 41.02000000000001 40.35 41.06000000000001 40.35 41.06000000000001 Z" stroke-linecap="round" />
-</g>
-<g transform="matrix(0.07 0 0 0.07 396.05 123.14)" style="" id="eb0df536-c517-4781-a7c0-3f84cd77c272"  >
-		<text xml:space="preserve" font-family="Lato" font-size="40" font-style="normal" font-weight="400" style="stroke: none; stroke-width: 1; stroke-dasharray: none; stroke-linecap: butt; stroke-dashoffset: 0; stroke-linejoin: miter; stroke-miterlimit: 4; fill: rgb(0,0,0); fill-rule: nonzero; opacity: 1; white-space: pre;" ><tspan x="-130" y="12.57" >Read The Docs</tspan></text>
-</g>
-<g transform="matrix(0.28 0 0 0.28 27.88 36)" id="7b9eddb9-1652-4040-9437-2ab90652d624"  >
-<path style="stroke: rgb(0,0,0); stroke-width: 0; stroke-dasharray: none; stroke-linecap: butt; stroke-dashoffset: 0; stroke-linejoin: miter; stroke-miterlimit: 4; fill: rgb(50,50,42); fill-rule: nonzero; opacity: 1;" vector-effect="non-scaling-stroke"  transform=" translate(-64, -64)" d="M 57.62 61.68 C 55.919999999999995 61.92 54.75 63.46 55 65.11 C 55.1668510745875 66.32380621250819 56.039448735907676 67.32218371690155 57.22 67.65 C 57.22 67.65 64.69 70.11 77.4 71.16000000000001 C 87.61000000000001 72.01 99.2 70.43 99.2 70.43 C 100.9 70.39 102.23 68.98 102.19 67.28 C 102.17037752125772 66.4652516996782 101.82707564255573 65.69186585376654 101.23597809465886 65.13079230830253 C 100.644880546762 64.56971876283853 99.85466451370849 64.26716220997277 99.03999999999999 64.29 C 98.83999999999999 64.29 98.63999999999999 64.33000000000001 98.42999999999999 64.37 C 98.42999999999999 64.37 87.08999999999999 65.78 77.88 65.02000000000001 C 65.72999999999999 64.05000000000001 59.11 61.83000000000001 59.11 61.83000000000001 C 58.63 61.670000000000016 58.1 61.59000000000001 57.62 61.670000000000016 Z M 57.62 46.46 C 55.919999999999995 46.7 54.75 48.24 55 49.89 C 55.1668510745875 51.10380621250818 56.039448735907676 52.10218371690154 57.22 52.43 C 57.22 52.43 64.69 54.89 77.4 55.94 C 87.61000000000001 56.79 99.2 55.21 99.2 55.21 C 100.9 55.17 102.23 53.76 102.19 52.06 C 102.17037752125772 51.245251699678214 101.82707564255573 50.47186585376654 101.23597809465886 49.91079230830253 C 100.644880546762 49.34971876283853 99.85466451370849 49.047162209972754 99.03999999999999 49.07 C 98.83999999999999 49.07 98.63999999999999 49.11 98.42999999999999 49.15 C 98.42999999999999 49.15 87.08999999999999 50.559999999999995 77.88 49.8 C 65.72999999999999 48.83 59.11 46.61 59.11 46.61 C 58.63 46.45 58.1 46.37 57.62 46.45 Z M 57.62 31.240000000000002 C 55.919999999999995 31.48 54.75 33.02 55 34.67 C 55.1668510745875 35.88380621250818 56.039448735907676 36.882183716901544 57.22 37.21 C 57.22 37.21 64.69 39.67 77.4 40.72 C 87.61000000000001 41.57 99.2 39.99 99.2 39.99 C 100.9 39.95 102.23 38.54 102.19 36.84 C 102.17037752125772 36.025251699678215 101.82707564255573 35.25186585376654 101.23597809465886 34.690792308302534 C 100.644880546762 34.12971876283853 99.85466451370849 33.827162209972755 99.03999999999999 33.85 C 98.83999999999999 33.85 98.63999999999999 33.89 98.42999999999999 33.93 C 98.42999999999999 33.93 87.08999999999999 35.339999999999996 77.88 34.58 C 65.72999999999999 33.61 59.11 31.389999999999997 59.11 31.389999999999997 C 58.63 31.229999999999997 58.1 31.189999999999998 57.62 31.229999999999997 Z M 57.62 16.060000000000002 C 55.919999999999995 16.3 54.75 17.840000000000003 55 19.490000000000002 C 55.1668510745875 20.703806212508187 56.039448735907676 21.702183716901544 57.22 22.03 C 57.22 22.03 64.69 24.490000000000002 77.4 25.54 C 87.61000000000001 26.39 99.2 24.81 99.2 24.81 C 100.9 24.77 102.23 23.36 102.19 21.66 C 102.17037752125772 20.84525169967821 101.82707564255573 20.07186585376654 101.23597809465886 19.510792308302534 C 100.644880546762 18.949718762838526 99.8546645137085 18.64716220997276 99.03999999999999 18.67 C 98.83999999999999 18.67 98.63999999999999 18.71 98.42999999999999 18.75 C 98.42999999999999 18.75 87.08999999999999 20.16 77.88 19.4 C 65.72999999999999 18.43 59.11 16.209999999999997 59.11 16.209999999999997 C 58.637850878541954 16.01924514007714 58.12188500879498 15.963839409097599 57.62 16.049999999999997 Z M 36.31 0 C 20.32 0.12 14.39 5.05 14.39 5.05 L 14.39 124.42 C 14.39 124.42 20.2 119.41 38.93 120.18 C 57.66 120.95000000000002 61.5 127.53 84.50999999999999 127.97000000000001 C 107.52 128.41000000000003 113.28999999999999 124.42000000000002 113.28999999999999 124.42000000000002 L 113.60999999999999 2.750000000000014 C 113.60999999999999 2.750000000000014 103.28 5.7 83.09 5.86 C 62.95 6.01 58.11 0.73 39.62 0.12 C 38.49 0.04 37.4 0 36.31 0 Z M 49.67 7.79 C 49.67 7.79 59.36 10.98 77.24000000000001 11.870000000000001 C 92.38000000000001 12.64 107.52000000000001 10.38 107.52000000000001 10.38 L 107.52000000000001 118.53 C 107.52000000000001 118.53 99.85000000000001 122.57000000000001 80.68 121.19 C 65.82000000000001 120.14 49.480000000000004 114.49 49.480000000000004 114.49 L 49.68000000000001 7.799999999999997 Z M 40.35 10.620000000000001 C 42.050000000000004 10.620000000000001 43.46 11.990000000000002 43.46 13.73 C 43.46 15.469999999999999 42.09 16.84 40.35 16.84 C 40.35 16.84 35.34 16.88 32.28 17.16 C 27.150000000000002 17.68 23.64 19.54 23.64 19.54 C 22.150000000000002 20.349999999999998 20.25 19.74 19.48 18.25 C 18.67 16.76 19.28 14.86 20.77 14.09 C 22.259999999999998 13.32 25.33 11.67 31.67 11.06 C 35.34 10.66 40.35 10.620000000000001 40.35 10.620000000000001 Z M 37.36 25.880000000000003 C 39.06 25.840000000000003 40.35 25.880000000000003 40.35 25.880000000000003 C 42.050000000000004 26.080000000000002 43.260000000000005 27.62 43.050000000000004 29.310000000000002 C 42.88374644848126 30.726609090871516 41.76660909087151 31.843746448481262 40.35 32.010000000000005 C 40.35 32.010000000000005 35.34 32.050000000000004 32.28 32.330000000000005 C 27.150000000000002 32.85000000000001 23.64 34.71000000000001 23.64 34.71000000000001 C 22.150000000000002 35.52000000000001 20.25 34.91000000000001 19.48 33.42000000000001 C 18.67 31.93000000000001 19.28 30.03000000000001 20.77 29.26000000000001 C 20.77 29.26000000000001 25.33 26.84000000000001 31.67 26.230000000000008 C 33.53 25.99000000000001 35.67 25.910000000000007 37.36 25.870000000000008 Z M 40.35 41.06 C 42.050000000000004 41.06 43.46 42.43 43.46 44.17 C 43.46 45.910000000000004 42.09 47.28 40.35 47.28 C 40.35 47.28 35.34 47.24 32.28 47.56 C 27.150000000000002 48.080000000000005 23.64 49.940000000000005 23.64 49.940000000000005 C 22.150000000000002 50.75000000000001 20.25 50.14000000000001 19.48 48.650000000000006 C 18.67 47.160000000000004 19.28 45.260000000000005 20.77 44.49000000000001 C 20.77 44.49000000000001 25.33 42.07000000000001 31.67 41.46000000000001 C 35.34 41.02000000000001 40.35 41.06000000000001 40.35 41.06000000000001 Z" stroke-linecap="round" />
-</g>
-<g transform="matrix(0.9 0 0 0.9 94 36)" style="" id="385bde16-f9fa-4222-bfea-1d5d5efcf730"  >
-		<text xml:space="preserve" font-family="Lato" font-size="15" font-style="normal" font-weight="100" style="stroke: none; stroke-width: 1; stroke-dasharray: none; stroke-linecap: butt; stroke-dashoffset: 0; stroke-linejoin: miter; stroke-miterlimit: 4; fill: rgb(0,0,0); fill-rule: nonzero; opacity: 1; white-space: pre;" ><tspan x="-48.68" y="4.71" >Read The Docs</tspan></text>
-</g>
-</svg>
\ No newline at end of file