PDF识别提取专家

PaddleOCR Chinese Document OCR — Reference#

Scanned Chinese document OCR with PP-OCRv4. ~98% simplified Chinese accuracy. Supports traditional Chinese, vertical text, and handwriting.

Environment#

• Venv: ~/.venvs/paddleocr/
• Python: ~/.venvs/paddleocr/bin/python
• Dependencies: paddlepaddle==2.6.2 + paddleocr==2.7.3 + numpy==1.26.4 (numpy<2 avoids opencv ABI conflict)

Installation#

python3 -m venv ~/.venvs/paddleocr
~/.venvs/paddleocr/bin/pip install paddlepaddle==2.6.2
~/.venvs/paddleocr/bin/pip install "numpy<2" paddleocr==2.7.3

Python 3.13 Compatibility (Important!)#

⚠️ PaddleOCR has multiple compatibility issues with Python 3.13. You MUST fix them manually.

Issue 1: imghdr Module Removed#

Symptom:

ModuleNotFoundError: No module named 'imghdr'

Cause: Python 3.13 removed the imghdr module (originally used for image type detection).

Solution: Create a shim module:

# Save to site-packages/imghdr.py
"""imghdr module shim for Python 3.13+ compatibility"""
def whatfile(f):
    return None

def what(buf, h=None):
    return None

Issue 2: np.sctypes Removed#

Symptom:

AttributeError: module 'numpy' has no attribute 'sctypes'

Cause: NumPy 2.0 removed np.sctypes attribute (used by imgaug package).

Solution: Edit site-packages/imgaug/imgaug.py lines 44-46:

# Original code (BROKEN):
# NP_FLOAT_TYPES = set(np.sctypes["float"])
# NP_INT_TYPES = set(np.sctypes["int"])
# NP_UINT_TYPES = set(np.sctypes["uint"])

# Fixed code:
NP_FLOAT_TYPES = {np.float16, np.float32, np.float64}
NP_INT_TYPES = {np.int8, np.int16, np.int32, np.int64}
NP_UINT_TYPES = {np.uint8, np.uint16, np.uint32, np.uint64}

Issue 3: OpenCV Incompatible with NumPy 2.x#

Symptom:

ImportError: numpy.core.multiarray failed to import

Cause: opencv-python 4.6 was compiled against NumPy 1.x, incompatible with NumPy 2.x ABI.

Solution: Use opencv-python-headless 4.13.0.92+:

~/.venvs/paddleocr/bin/pip uninstall opencv-python -y
~/.venvs/paddleocr/bin/pip install opencv-python-headless --no-cache-dir

Issue 4: scikit-image DLL Load Failed (Windows)#

Symptom:

ImportError: DLL load failed while importing _cython_blas

Cause: Windows security policy blocks unsigned C extension DLLs.

Solutions:

• Option A: Use pre-compiled wheel (Recommended)
• Option B: Temporarily disable Windows Defender Real-time Protection
• Option C: Use WSL2 or Linux environment

Recommended Python 3.13 Installation#

# Environment
python3 -m venv ~/.venvs/paddleocr
source ~/.venvs/paddleocr/bin/activate

# Install (use PaddlePaddle 3.0.0 which supports NumPy 2.x)
pip install paddlepaddle==3.0.0
pip install "numpy>=2" paddleocr==2.7.3
pip uninstall opencv-python -y
pip install opencv-python-headless --no-cache-dir

# Create imghdr shim
cat > ~/.venvs/paddleocr/lib/python3.13/site-packages/imghdr.py << 'EOF'
def whatfile(f):
    return None
def what(buf, h=None):
    return None
EOF

# Patch np.sctypes issue
# Manually edit ~/.venvs/paddleocr/lib/python3.13/site-packages/imgaug/imgaug.py

---

CLI: OCR a Single Image#

~/.venvs/paddleocr/bin/python -c "
from paddleocr import PaddleOCR
ocr = PaddleOCR(lang='ch')  # auto-downloads models on first run
result = ocr.ocr('image.png')
for line_info in result[0]:
    text = line_info[1][0]
    conf = line_info[1][1]
    print(f'{conf:.2f} {text}')
"

Workflow: Scanned PDF → Structured Markdown#

Step 1: Split pages#

Preferred (if OpenDataLoader installed):

opendataloader-pdf input_scan.pdf -f text -o ./pages

Alternative (pdf2image):

pip install pdf2image
python -c "
from pdf2image import convert_from_path
images = convert_from_path('input.pdf', dpi=300)
for i, img in enumerate(images):
    img.save(f'page_{i+1:03d}.png')
"

Step 2: OCR each page#

from paddleocr import PaddleOCR
import os

ocr = PaddleOCR(lang='ch', use_angle_cls=True, use_gpu=False)
output_pages = {}

png_files = sorted([f for f in os.listdir('./pages') if f.endswith('.png')],
                   key=lambda x: int(''.join(filter(str.isdigit, x)) or 0))

for png in png_files:
    result = ocr.ocr(f'./pages/{png}')
    page_num = ''.join(filter(str.isdigit, png))
    lines = []
    for line_info in result[0]:
        text = line_info[1][0]
        conf = line_info[1][1]
        if conf > 0.6:
            lines.append(text)
    output_pages[page_num] = lines

Step 3: Assemble Markdown#

# OCR Result — filename
> Engine: PaddleOCR PP-OCRv4 (Simplified Chinese)
> Pages: 17 | Date: YYYY-MM-DD

---

## Page 1

Body text...

---

## Page 2

Body text...

Accuracy#

Engine	Simplified Chinese	Traditional Chinese	Seals/Signatures	Vertical Text
Tesseract chi_sim	~80%	~60%	❌ garbled	❌ not supported
PaddleOCR PP-OCRv4	~98%	~95%	✅ mostly correct	✅ supported
PaddleOCR + 预处理	~98%+	~96%	✅	✅

预处理（倾斜校正/文档展平）对手机拍照、弯曲文档提升 1~5%。 详见 SKILL.md D 节 scripts/pdf_preprocess.py。

Troubleshooting#

Poor character accuracy?#

• Increase DPI: minimum 300, recommended 400-600
• Lower confidence threshold: if conf > 0.3 to retain more candidates
• Verify critical pages (signatures, amounts) with vision_analyze

Poor table recognition?#

PaddleOCR's table recognition (lang='ch' + use_angle_cls=True) is moderate. For dense tables, combine with the main pdf-ocr-md docling+PaddleOCR pipeline.

GPU acceleration#

~/.venvs/paddleocr/bin/pip install paddlepaddle-gpu==2.6.2

Set use_gpu=True when creating PaddleOCR.