refactor(ocr):Increase the dilation factor in OCR to address the issue of word concatenation.

- Remove unused functions such as split_long_words, ocr_mk_mm_markdown_with_para, etc.
- Simplify ocr_mk_markdown_with_para_core_v2 by removing unnecessary language detection and word splitting logic- Remove wordninja dependency from requirements
- Update ocr_model_init to include additional parameters for OCR model configuration

magic_pdf/dict2md/ocr_mkcontent.py

@@ -1,6 +1,5 @@
 import re
 
-import wordninja
 from loguru import logger
 
 from magic_pdf.libs.commons import join_path
@@ -25,37 +24,6 @@ def __is_hyphen_at_line_end(line):
     return bool(re.search(r'[A-Za-z]+-\s*$', line))
 
 
-def split_long_words(text):
-    segments = text.split(' ')
-    for i in range(len(segments)):
-        words = re.findall(r'\w+|[^\w]', segments[i], re.UNICODE)
-        for j in range(len(words)):
-            if len(words[j]) > 10:
-                words[j] = ' '.join(wordninja.split(words[j]))
-        segments[i] = ''.join(words)
-    return ' '.join(segments)
-
-
-def ocr_mk_mm_markdown_with_para(pdf_info_list: list, img_buket_path):
-    markdown = []
-    for page_info in pdf_info_list:
-        paras_of_layout = page_info.get('para_blocks')
-        page_markdown = ocr_mk_markdown_with_para_core_v2(
-            paras_of_layout, 'mm', img_buket_path)
-        markdown.extend(page_markdown)
-    return '\n\n'.join(markdown)
-
-
-def ocr_mk_nlp_markdown_with_para(pdf_info_dict: list):
-    markdown = []
-    for page_info in pdf_info_dict:
-        paras_of_layout = page_info.get('para_blocks')
-        page_markdown = ocr_mk_markdown_with_para_core_v2(
-            paras_of_layout, 'nlp')
-        markdown.extend(page_markdown)
-    return '\n\n'.join(markdown)
-
-
 def ocr_mk_mm_markdown_with_para_and_pagination(pdf_info_dict: list,
                                                 img_buket_path):
     markdown_with_para_and_pagination = []
@@ -76,45 +44,6 @@ def ocr_mk_mm_markdown_with_para_and_pagination(pdf_info_dict: list,
     return markdown_with_para_and_pagination
 
 
-def ocr_mk_markdown_with_para_core(paras_of_layout, mode, img_buket_path=''):
-    page_markdown = []
-    for paras in paras_of_layout:
-        for para in paras:
-            para_text = ''
-            for line in para:
-                for span in line['spans']:
-                    span_type = span.get('type')
-                    content = ''
-                    language = ''
-                    if span_type == ContentType.Text:
-                        content = span['content']
-                        language = detect_lang(content)
-                        if (language == 'en'):  # 只对英文长词进行分词处理，中文分词会丢失文本
-                            content = ocr_escape_special_markdown_char(
-                                split_long_words(content))
-                        else:
-                            content = ocr_escape_special_markdown_char(content)
-                    elif span_type == ContentType.InlineEquation:
-                        content = f"${span['content']}$"
-                    elif span_type == ContentType.InterlineEquation:
-                        content = f"\n$$\n{span['content']}\n$$\n"
-                    elif span_type in [ContentType.Image, ContentType.Table]:
-                        if mode == 'mm':
-                            content = f"\n![]({join_path(img_buket_path, span['image_path'])})\n"
-                        elif mode == 'nlp':
-                            pass
-                    if content != '':
-                        if language == 'en':  # 英文语境下 content间需要空格分隔
-                            para_text += content + ' '
-                        else:  # 中文语境下，content间不需要空格分隔
-                            para_text += content
-            if para_text.strip() == '':
-                continue
-            else:
-                page_markdown.append(para_text.strip() + '  ')
-    return page_markdown
-
-
 def ocr_mk_markdown_with_para_core_v2(paras_of_layout,
                                       mode,
                                       img_buket_path='',
@@ -207,21 +136,11 @@ def merge_para_with_text(para_block, parse_type="auto", lang=None):
         if line_text != '':
             line_lang = detect_lang(line_text)
         for span in line['spans']:
+
             span_type = span['type']
             content = ''
             if span_type == ContentType.Text:
-                content = span['content']
-                # language = detect_lang(content)
-                language = detect_language(content)
-                # 判断是否小语种
-                if lang is not None and lang != 'en':
-                    content = ocr_escape_special_markdown_char(content)
-                else:  # 非小语种逻辑
-                    if language == 'en' and parse_type == 'ocr':  # 只对英文长词进行分词处理，中文分词会丢失文本
-                        content = ocr_escape_special_markdown_char(
-                            split_long_words(content))
-                    else:
-                        content = ocr_escape_special_markdown_char(content)
+                content = ocr_escape_special_markdown_char(span['content'])
             elif span_type == ContentType.InlineEquation:
                 content = f" ${span['content']}$ "
             elif span_type == ContentType.InterlineEquation:
@@ -242,41 +161,6 @@ def merge_para_with_text(para_block, parse_type="auto", lang=None):
     return para_text
 
 
-def para_to_standard_format(para, img_buket_path):
-    para_content = {}
-    if len(para) == 1:
-        para_content = line_to_standard_format(para[0], img_buket_path)
-    elif len(para) > 1:
-        para_text = ''
-        inline_equation_num = 0
-        for line in para:
-            for span in line['spans']:
-                language = ''
-                span_type = span.get('type')
-                content = ''
-                if span_type == ContentType.Text:
-                    content = span['content']
-                    language = detect_lang(content)
-                    if language == 'en':  # 只对英文长词进行分词处理，中文分词会丢失文本
-                        content = ocr_escape_special_markdown_char(
-                            split_long_words(content))
-                    else:
-                        content = ocr_escape_special_markdown_char(content)
-                elif span_type == ContentType.InlineEquation:
-                    content = f"${span['content']}$"
-                    inline_equation_num += 1
-                if language == 'en':  # 英文语境下 content间需要空格分隔
-                    para_text += content + ' '
-                else:  # 中文语境下，content间不需要空格分隔
-                    para_text += content
-        para_content = {
-            'type': 'text',
-            'text': para_text,
-            'inline_equation_num': inline_equation_num,
-        }
-    return para_content
-
-
 def para_to_standard_format_v2(para_block, img_buket_path, page_idx, parse_type="auto", lang=None, drop_reason=None):
     para_type = para_block['type']
     para_content = {}
@@ -330,82 +214,6 @@ def para_to_standard_format_v2(para_block, img_buket_path, page_idx, parse_type=
     return para_content
 
 
-def make_standard_format_with_para(pdf_info_dict: list, img_buket_path: str):
-    content_list = []
-    for page_info in pdf_info_dict:
-        paras_of_layout = page_info.get('para_blocks')
-        if not paras_of_layout:
-            continue
-        for para_block in paras_of_layout:
-            para_content = para_to_standard_format_v2(para_block,
-                                                      img_buket_path)
-            content_list.append(para_content)
-    return content_list
-
-
-def line_to_standard_format(line, img_buket_path):
-    line_text = ''
-    inline_equation_num = 0
-    for span in line['spans']:
-        if not span.get('content'):
-            if not span.get('image_path'):
-                continue
-            else:
-                if span['type'] == ContentType.Image:
-                    content = {
-                        'type': 'image',
-                        'img_path': join_path(img_buket_path,
-                                              span['image_path']),
-                    }
-                    return content
-                elif span['type'] == ContentType.Table:
-                    content = {
-                        'type': 'table',
-                        'img_path': join_path(img_buket_path,
-                                              span['image_path']),
-                    }
-                    return content
-        else:
-            if span['type'] == ContentType.InterlineEquation:
-                interline_equation = span['content']
-                content = {
-                    'type': 'equation',
-                    'latex': f'$$\n{interline_equation}\n$$'
-                }
-                return content
-            elif span['type'] == ContentType.InlineEquation:
-                inline_equation = span['content']
-                line_text += f'${inline_equation}$'
-                inline_equation_num += 1
-            elif span['type'] == ContentType.Text:
-                text_content = ocr_escape_special_markdown_char(
-                    span['content'])  # 转义特殊符号
-                line_text += text_content
-    content = {
-        'type': 'text',
-        'text': line_text,
-        'inline_equation_num': inline_equation_num,
-    }
-    return content
-
-
-def ocr_mk_mm_standard_format(pdf_info_dict: list):
-    """content_list type         string
-    image/text/table/equation(行间的单独拿出来，行内的和text合并) latex        string
-    latex文本字段。 text         string      纯文本格式的文本数据。 md           string
-    markdown格式的文本数据。 img_path     string      s3://full/path/to/img.jpg."""
-    content_list = []
-    for page_info in pdf_info_dict:
-        blocks = page_info.get('preproc_blocks')
-        if not blocks:
-            continue
-        for block in blocks:
-            for line in block['lines']:
-                content = line_to_standard_format(line)
-                content_list.append(content)
-    return content_list
-
-
 def union_make(pdf_info_dict: list,
                make_mode: str,
                drop_mode: str,

magic_pdf/model/pdf_extract_kit.py

@@ -77,11 +77,11 @@ def layout_model_init(weight, config_file, device):
     return model
 
 
-def ocr_model_init(show_log: bool = False, det_db_box_thresh=0.3, lang=None):
+def ocr_model_init(show_log: bool = False, det_db_box_thresh=0.3, lang=None, use_dilation=True, det_db_unclip_ratio=2.4):
     if lang is not None:
-        model = ModifiedPaddleOCR(show_log=show_log, det_db_box_thresh=det_db_box_thresh, lang=lang)
+        model = ModifiedPaddleOCR(show_log=show_log, det_db_box_thresh=det_db_box_thresh, lang=lang, use_dilation=use_dilation, det_db_unclip_ratio=det_db_unclip_ratio)
     else:
-        model = ModifiedPaddleOCR(show_log=show_log, det_db_box_thresh=det_db_box_thresh)
+        model = ModifiedPaddleOCR(show_log=show_log, det_db_box_thresh=det_db_box_thresh, use_dilation=use_dilation, det_db_unclip_ratio=det_db_unclip_ratio)
     return model
 
 

requirements-docker.txt

@@ -5,7 +5,6 @@ PyMuPDF>=1.24.9
 loguru>=0.6.0
 numpy>=1.21.6,<2.0.0
 fast-langdetect==0.2.0
-wordninja>=2.0.0
 scikit-learn>=1.0.2
 pdfminer.six==20231228
 unimernet==0.2.1

requirements.txt

@@ -8,7 +8,6 @@ pdfminer.six==20231228
 pydantic>=2.7.2,<2.8.0
 PyMuPDF>=1.24.9
 scikit-learn>=1.0.2
-wordninja>=2.0.0
 torch>=2.2.2,<=2.3.1
 transformers
 # The requirements.txt must ensure that only necessary external dependencies are introduced. If there are new dependencies to add, please contact the project administrator.