opensource/MinerU
3
0
Fork 0
mirror of https://github.com/opendatalab/MinerU.git synced 2026-09-24 23:10:23 +08:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #5073 from myhloli/dev

Browse Source 
Refactor OCR functions and enhance text rendering metrics
This commit is contained in:
Xiaomeng Zhao
2026-06-04 11:11:19 +08:00
committed by GitHub
parent ecf23267f1 751aae3b4a
commit 52d77c6ed5
7 changed files with 256 additions and 37 deletions
Download Patch File Download Diff File Expand all files Collapse all files
mineru/backend/hybrid/hybrid_analyze.py
+4 -5
View File
@@ -23,8 +23,7 @@ from mineru.backend.pipeline.model_init import (
HybridModelSingleton,
run_layout_inference,
run_mfr_inference,
run_ocr_det_inference,
run_ocr_rec_inference,
run_ocr_inference,
)
from mineru.backend.vlm.vlm_analyze import (
ModelSingleton,
@@ -125,7 +124,7 @@ def ocr_det(
page_mfd_res, useful_list
)
bgr_image = cv2.cvtColor(new_image, cv2.COLOR_RGB2BGR)
ocr_res = run_ocr_det_inference(
ocr_res = run_ocr_inference(
hybrid_pipeline_model.ocr_model.ocr,
bgr_image,
mfd_res=adjusted_mfdetrec_res,
@@ -202,7 +201,7 @@ def ocr_det(
# 批处理检测
det_batch_size = min(len(batch_images), batch_ratio * OCR_DET_BASE_BATCH_SIZE)
batch_results = run_ocr_det_inference(
batch_results = run_ocr_inference(
hybrid_pipeline_model.ocr_model.text_detector.batch_predict,
batch_images,
det_batch_size,
@@ -488,7 +487,7 @@ def _process_ocr_and_formulas(
img_crop_list.append(ocr_res.pop('np_img'))
if len(img_crop_list) > 0:
# Process OCR
ocr_result_list = run_ocr_rec_inference(
ocr_result_list = run_ocr_inference(
hybrid_pipeline_model.ocr_model.ocr,
img_crop_list,
det=False,
mineru/backend/hybrid/hybrid_model_output_to_middle_json.py
+9 -2
View File
@@ -14,12 +14,16 @@ from mineru.backend.utils.para_block_utils import (
)
from mineru.backend.hybrid.hybrid_magic_model import MagicModel
from mineru.backend.utils.runtime_utils import cross_page_table_merge
from mineru.backend.pipeline.model_init import run_ocr_rec_inference
from mineru.backend.pipeline.model_init import run_ocr_inference
from mineru.utils.config_reader import get_table_enable
from mineru.utils.cut_image import cut_image_and_table
from mineru.utils.enum_class import ContentType, BlockType
from mineru.utils.hash_utils import bytes_md5
from mineru.utils.ocr_utils import OcrConfidence, rotate_vertical_crop_if_needed
from mineru.utils.span_pre_proc import (
_clear_post_ocr_fallback,
_restore_post_ocr_fallback,
)
from mineru.utils.title_level_postprocess import apply_title_leveling_to_pdf_info
from mineru.utils.pdfium_guard import close_pdfium_child, close_pdfium_document, pdfium_guard
from mineru.version import __version__
@@ -139,7 +143,7 @@ def _apply_post_ocr(pdf_info_list, hybrid_pipeline_model):
img_crop_list.append(rotate_vertical_crop_if_needed(span['np_img']))
span.pop('np_img')
if len(img_crop_list) > 0:
ocr_res_list = run_ocr_rec_inference(
ocr_res_list = run_ocr_inference(
hybrid_pipeline_model.ocr_model.ocr,
img_crop_list,
det=False,
@@ -152,6 +156,9 @@ def _apply_post_ocr(pdf_info_list, hybrid_pipeline_model):
if ocr_score > OcrConfidence.min_confidence:
span['content'] = ocr_text
span['score'] = float(f"{ocr_score:.3f}")
_clear_post_ocr_fallback(span)
elif _restore_post_ocr_fallback(span):
continue
else:
span['content'] = ''
span['score'] = 0.0
mineru/backend/pipeline/batch_analyze.py
+7 -8
View File
@@ -13,8 +13,7 @@ from .model_init import (
AtomModelSingleton,
run_layout_inference,
run_mfr_inference,
run_ocr_det_inference,
run_ocr_rec_inference,
run_ocr_inference,
)
from .model_list import AtomicModel
from ...utils.config_reader import (
@@ -534,7 +533,7 @@ class BatchAnalyze:
if inline_mask_boxes
else bgr_image
)
ocr_result = run_ocr_det_inference(
ocr_result = run_ocr_inference(
det_ocr_engine.ocr, det_image, rec=False
)[0]
if ocr_result and formula_mask_boxes:
@@ -565,7 +564,7 @@ class BatchAnalyze:
enable_merge_det_boxes=False,
)
cropped_img_list = [item["cropped_img"] for item in rec_img_list]
ocr_res_list = run_ocr_rec_inference(
ocr_res_list = run_ocr_inference(
ocr_engine.ocr,
cropped_img_list,
det=False,
@@ -731,7 +730,7 @@ class BatchAnalyze:
# 批处理检测
det_batch_size = min(len(batch_images), self.batch_ratio * OCR_DET_BASE_BATCH_SIZE)
batch_results = run_ocr_det_inference(
batch_results = run_ocr_inference(
ocr_model.text_detector.batch_predict, batch_images, det_batch_size
)
@@ -793,7 +792,7 @@ class BatchAnalyze:
bgr_image,
adjusted_mfdetrec_res,
)
ocr_res = run_ocr_det_inference(
ocr_res = run_ocr_inference(
ocr_model.ocr,
det_image,
mfd_res=adjusted_mfdetrec_res,
@@ -852,7 +851,7 @@ class BatchAnalyze:
atom_model_name=AtomicModel.OCR,
lang=lang
)
ocr_res_list = run_ocr_rec_inference(
ocr_res_list = run_ocr_inference(
ocr_model.ocr, img_crop_list, det=False, tqdm_enable=True
)[0]
@@ -923,7 +922,7 @@ class BatchAnalyze:
)
seal_crop_bgr = cv2.cvtColor(seal_crop_rgb, cv2.COLOR_RGB2BGR)
seal_ocr_res = run_ocr_det_inference(
seal_ocr_res = run_ocr_inference(
seal_ocr_model.ocr, seal_crop_bgr, det=True, rec=True
)[0]
if not seal_ocr_res:
mineru/backend/pipeline/model_init.py
+4 -12
View File
@@ -22,8 +22,7 @@ PIPELINE_MODEL_INIT_LOCK = threading.RLock()
# 这些锁保护 pipeline 与 hybrid 共享的 atom model/native 模型推理调用，避免多线程同时进入同一个模型对象。
PIPELINE_LAYOUT_INFERENCE_LOCK = threading.RLock()
PIPELINE_MFR_INFERENCE_LOCK = threading.RLock()
PIPELINE_OCR_DET_INFERENCE_LOCK = threading.RLock()
PIPELINE_OCR_REC_INFERENCE_LOCK = threading.RLock()
PIPELINE_OCR_INFERENCE_LOCK = threading.RLock()
# 临时关闭 pipeline/hybrid 共享推理阶段锁；需要回滚实验时可通过环境变量重新打开。
PIPELINE_INFERENCE_LOCKS_ENABLED = os.getenv(
'MINERU_ENABLE_PIPELINE_INFERENCE_LOCKS', 'False'
@@ -53,17 +52,10 @@ def run_mfr_inference(inference_callable, *args, **kwargs):
)
def run_ocr_det_inference(inference_callable, *args, **kwargs):
"""按实验开关执行共享 OCR det 模型调用。"""
def run_ocr_inference(inference_callable, *args, **kwargs):
"""按实验开关执行共享 OCR native 模型调用。"""
return _run_with_inference_lock(
PIPELINE_OCR_DET_INFERENCE_LOCK, inference_callable, *args, **kwargs
)
def run_ocr_rec_inference(inference_callable, *args, **kwargs):
"""按实验开关执行共享 OCR rec 模型调用。"""
return _run_with_inference_lock(
PIPELINE_OCR_REC_INFERENCE_LOCK, inference_callable, *args, **kwargs
PIPELINE_OCR_INFERENCE_LOCK, inference_callable, *args, **kwargs
)
MFR_MODEL = os.getenv('MINERU_FORMULA_CH_SUPPORT', 'False')
mineru/backend/pipeline/model_json_to_middle_json.py
+9 -2
View File
@@ -7,7 +7,7 @@ from mineru.backend.utils.html_image_utils import replace_inline_table_images
from mineru.backend.utils.runtime_utils import cross_page_table_merge
from mineru.backend.pipeline.model_init import (
AtomModelSingleton,
run_ocr_rec_inference,
run_ocr_inference,
)
from mineru.backend.pipeline.para_split import para_split
from mineru.utils.char_utils import full_to_half
@@ -19,6 +19,10 @@ from mineru.utils.ocr_utils import OcrConfidence, rotate_vertical_crop_if_needed
from mineru.version import __version__
from mineru.utils.hash_utils import bytes_md5
from mineru.utils.pdfium_guard import close_pdfium_child, close_pdfium_document, pdfium_guard
from mineru.utils.span_pre_proc import (
_clear_post_ocr_fallback,
_restore_post_ocr_fallback,
)
def page_model_info_to_page_info(page_model_info, image_dict, page, image_writer, page_index, ocr_enable=False):
@@ -235,7 +239,7 @@ def _apply_post_ocr(pdf_info_list, lang=None):
det_db_box_thresh=0.3,
lang=lang
)
ocr_res_list = run_ocr_rec_inference(
ocr_res_list = run_ocr_inference(
ocr_model.ocr, img_crop_list, det=False, tqdm_enable=True
)[0]
assert len(ocr_res_list) == len(
@@ -245,6 +249,9 @@ def _apply_post_ocr(pdf_info_list, lang=None):
if ocr_score > OcrConfidence.min_confidence:
span['content'] = ocr_text
span['score'] = float(f"{ocr_score:.3f}")
_clear_post_ocr_fallback(span)
elif _restore_post_ocr_fallback(span):
continue
else:
span['content'] = ''
span['score'] = 0.0
mineru/utils/span_pre_proc.py
+216 -8
View File
@@ -15,6 +15,15 @@ from mineru.utils.pdf_text_tool import get_lines_from_chars, get_page_chars
from mineru.utils.pdfium_guard import close_pdfium_child, pdfium_guard
MAX_NATIVE_TEXT_CHARS_PER_PAGE = 65535
PRIVATE_USE_AREA_START = 0xE000
PRIVATE_USE_AREA_END = 0xF8FF
PRIVATE_USE_TEXT_COUNT_THRESHOLD = 2
PRIVATE_USE_TEXT_RATIO_THRESHOLD = 0.05
PRIVATE_USE_TEXT_RUN_THRESHOLD = 2
POST_OCR_FALLBACK_CONTENT_KEY = '_post_ocr_fallback_content'
POST_OCR_FALLBACK_SCORE_KEY = '_post_ocr_fallback_score'
POST_OCR_REASON_KEY = '_post_ocr_reason'
POST_OCR_REASON_PRIVATE_USE_TEXT = 'private_use_text'
def __replace_ligatures(text: str):
@@ -141,6 +150,8 @@ def _prepare_post_ocr_spans(need_ocr_spans, spans, pil_img, scale):
span_img = cv2.cvtColor(np.array(span_pil_img), cv2.COLOR_RGB2BGR)
# 计算span的对比度，低于0.17的span不进行ocr，等于0.17的临界框保留给后置OCR。
if calculate_contrast(span_img, img_mode='bgr') < 0.17:
if _restore_post_ocr_fallback(span):
continue
if span in spans:
spans.remove(span)
continue
@@ -267,9 +278,18 @@ def fill_char_in_spans(spans, all_chars, median_span_height):
need_ocr_spans = []
for span in spans:
private_use_signal = _get_private_use_text_signal(span['chars'])
should_post_ocr_private_use = _should_fallback_to_post_ocr_for_private_use_text(
private_use_signal
)
chars_to_content(span)
if should_post_ocr_private_use and span.get('content'):
span[POST_OCR_FALLBACK_CONTENT_KEY] = span['content']
span[POST_OCR_FALLBACK_SCORE_KEY] = span.get('score', 1.0)
span[POST_OCR_REASON_KEY] = POST_OCR_REASON_PRIVATE_USE_TEXT
need_ocr_spans.append(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:
elif 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']
@@ -280,6 +300,83 @@ LINE_STOP_FLAG = ('.', '!', '?', '。', '！', '？', ')', '）', '"', '”', ':
LINE_START_FLAG = ('(', '（', '"', '“', '【', '{', '《', '<', '「', '『', '【', '[',)
Span_Height_Ratio = 0.33 # 字符的中轴和span的中轴高度差不能超过1/3span高度
SCRIPT_BODY_HEIGHT_RATIO = 0.9
SCRIPT_CENTER_TOLERANCE_RATIO = 0.12
def _is_private_use_char(char: str) -> bool:
"""判断单个字符是否落在 Unicode 私用区，用于识别字体映射异常。"""
return (
len(char) == 1
and PRIVATE_USE_AREA_START <= ord(char) <= PRIVATE_USE_AREA_END
)
def _get_private_use_text_signal(chars):
"""统计 span 字符中的私用区信号，供局部后置 OCR 决策使用。"""
pua_count = 0
text_char_count = 0
current_pua_run = 0
max_pua_run = 0
for char in chars:
for text_char in char.get('char', ''):
if text_char.isspace():
current_pua_run = 0
continue
text_char_count += 1
if _is_private_use_char(text_char):
pua_count += 1
current_pua_run += 1
max_pua_run = max(max_pua_run, current_pua_run)
else:
current_pua_run = 0
pua_ratio = 0.0
if text_char_count > 0:
pua_ratio = pua_count / text_char_count
return {
'pua_count': pua_count,
'text_char_count': text_char_count,
'pua_ratio': pua_ratio,
'max_pua_run': max_pua_run,
}
def _should_fallback_to_post_ocr_for_private_use_text(signal) -> bool:
"""连续或高占比 PUA 才转后置 OCR，降低孤立私用符号误召回。"""
pua_count = signal['pua_count']
if pua_count < PRIVATE_USE_TEXT_COUNT_THRESHOLD:
return False
return (
signal['max_pua_run'] >= PRIVATE_USE_TEXT_RUN_THRESHOLD
or signal['pua_ratio'] >= PRIVATE_USE_TEXT_RATIO_THRESHOLD
)
def _clear_post_ocr_fallback(span):
"""清理后置 OCR 内部兜底字段，避免进入最终 middle-json 输出。"""
span.pop(POST_OCR_FALLBACK_CONTENT_KEY, None)
span.pop(POST_OCR_FALLBACK_SCORE_KEY, None)
span.pop(POST_OCR_REASON_KEY, None)
def _restore_post_ocr_fallback(span) -> bool:
"""在后置 OCR 无法使用时恢复原始文本兜底，返回是否已恢复。"""
if POST_OCR_FALLBACK_CONTENT_KEY not in span:
_clear_post_ocr_fallback(span)
return False
span['content'] = span[POST_OCR_FALLBACK_CONTENT_KEY]
if POST_OCR_FALLBACK_SCORE_KEY in span:
span['score'] = span[POST_OCR_FALLBACK_SCORE_KEY]
_clear_post_ocr_fallback(span)
return True
def calculate_char_in_span(char_bbox, span_bbox, char, span_height_ratio=Span_Height_Ratio):
char_center_x = (char_bbox[0] + char_bbox[2]) / 2
char_center_y = (char_bbox[1] + char_bbox[3]) / 2
@@ -315,6 +412,114 @@ def calculate_char_in_span(char_bbox, span_bbox, char, span_height_ratio=Span_He
return False
def _get_char_bbox_metrics(char):
"""提取字符 bbox 的宽高和中心点，统一兼容 list 与 pdftext Bbox 对象。"""
bbox = char['bbox']
x0, y0, x1, y1 = [float(v) for v in bbox]
return {
'width': x1 - x0,
'height': y1 - y0,
'center_y': (y0 + y1) / 2,
}
def _get_char_bbox_metrics_list(chars):
"""预计算 span 内全部字符的 bbox 指标，避免上下标判断重复解析 bbox。"""
return [_get_char_bbox_metrics(char) for char in chars]
def _is_valid_script_reference_char(char, metrics) -> bool:
"""过滤空白和退化 bbox，只用真实可见字符估计正文主带。"""
if char['char'] in {' ', '\r', '\n'}:
return False
return metrics['height'] > 1 and metrics['width'] > 0
def _get_body_axis(chars, char_metrics):
"""根据同一 span 内最大高度字符簇估计正文中心线和正文高度。"""
valid_metrics = [
metrics for char, metrics in zip(chars, char_metrics)
if _is_valid_script_reference_char(char, metrics)
]
if not valid_metrics:
return None
max_height = max(metrics['height'] for metrics in valid_metrics)
body_metrics = [
metrics for metrics in valid_metrics
if metrics['height'] >= max_height * SCRIPT_BODY_HEIGHT_RATIO
]
if not body_metrics:
return None
return {
'center_y': statistics.median(metrics['center_y'] for metrics in body_metrics),
'height': statistics.median(metrics['height'] for metrics in body_metrics),
}
def _classify_char_script_roles(chars, char_metrics):
"""按正文主带判断每个字符属于正文、上标或下标。"""
body_axis = _get_body_axis(chars, char_metrics)
if body_axis is None or body_axis['height'] <= 0:
return ['body'] * len(chars)
tolerance = body_axis['height'] * SCRIPT_CENTER_TOLERANCE_RATIO
roles = []
for char, metrics in zip(chars, char_metrics):
if not _is_valid_script_reference_char(char, metrics):
roles.append('body')
continue
char_center_y = metrics['center_y']
if char_center_y < body_axis['center_y'] - tolerance:
roles.append('sup')
elif char_center_y > body_axis['center_y'] + tolerance:
roles.append('sub')
else:
roles.append('body')
return roles
def _append_script_wrapped_text(parts, role, text):
"""把连续同类上下标文本包裹成 HTML 标签，正文保持原样。"""
if not text:
return
if role == 'sup':
parts.append(f'<sup>{text}</sup>')
elif role == 'sub':
parts.append(f'<sub>{text}</sub>')
else:
parts.append(text)
def _wrap_script_runs(role_text_parts):
"""合并连续正文、上标、下标 run，避免每个字符单独生成标签。"""
wrapped_parts = []
current_role = None
current_text_parts = []
for role, text in role_text_parts:
if role != current_role:
_append_script_wrapped_text(
wrapped_parts,
current_role,
''.join(current_text_parts),
)
current_role = role
current_text_parts = [text]
else:
current_text_parts.append(text)
_append_script_wrapped_text(
wrapped_parts,
current_role,
''.join(current_text_parts),
)
return ''.join(wrapped_parts)
def chars_to_content(span):
# 检查span中的char是否为空
if len(span['chars']) != 0:
@@ -326,28 +531,31 @@ def chars_to_content(span):
):
chars = sorted(chars, key=lambda x: x['char_idx'])
char_metrics = _get_char_bbox_metrics_list(chars)
# Calculate the width of each character
char_widths = [char['bbox'][2] - char['bbox'][0] for char in chars]
char_widths = [metrics['width'] for metrics in char_metrics]
# Calculate the median width
median_width = statistics.median(char_widths)
script_roles = _classify_char_script_roles(chars, char_metrics)
parts = []
role_text_parts = []
for idx, char1 in enumerate(chars):
char2 = chars[idx + 1] if idx + 1 < len(chars) else None
role1 = script_roles[idx]
role2 = script_roles[idx + 1] if char2 else None
# 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
role_text_parts.append((role1, char1['char']))
if (
char2
and char2['bbox'][0] - char1['bbox'][2] > median_width * 0.25
and char1['char'] != ' '
and char2['char'] != ' '
):
parts.append(char1['char'])
parts.append(' ')
else:
parts.append(char1['char'])
space_role = role1 if role1 == role2 else 'body'
role_text_parts.append((space_role, ' '))
content = ''.join(parts)
content = _wrap_script_runs(role_text_parts)
content = __replace_unicode(content)
content = __replace_ligatures(content)
span['content'] = content.strip()
pyproject.toml
+7
View File
@@ -180,3 +180,10 @@ exclude_also = [
'class .*\bProtocol\):',
'@(abc\.)?abstractmethod',
]
[tool.ruff]
line-length = 128
[tool.ruff.lint]
select = ["C", "E", "F", "W", "ANN"]
ignore = ["C901", "ANN204", "ANN401"]