PaddleOCR-VL Recognition

This is a PaddleOCR-VL recognition script that retains only the core functionality of the VL recognition model from PaddleOCR.

Additionally, it has been adapted for fine-tuned PaddleOCR-VL models from AI Studio or from ModelScope , allowing recognition using custom prompt_label and query parameters.

Recognition Results Display

Input Image	Full Information Extraction	Specific Information Extraction

Environment Setup

Install Dependencies Using requirements.txt

pip install -r requirements.txt

Manual Installation (If Needed)

python -m pip install -U "paddleocr[doc-parser]"
python -m pip install https://paddle-whl.bj.bcebos.com/nightly/cu126/safetensors/safetensors-0.6.2.dev0-cp38-abi3-linux_x86_64.whl
python -m pip install --force-reinstall opencv-python-headless
python -m pip install numpy==1.26.4
python -m pip install json-repair

Features

1. ✅ Able to initialize vl_rec_model
2. ✅ Able to input an image and directly output recognized text results
3. ❌ Does not include document preprocessing (use_doc_preprocessor)
4. ❌ Does not include layout detection (use_layout_detection)
5. ❌ Does not include chart recognition (use_chart_recognition)
6. ❌ Does not include content formatting (format_block_content)
7. ❌ Does not include layout block merging (merge_layout_blocks)
8. ❌ Does not include markdown ignore labels (markdown_ignore_labels)
9. ❌ Does not output extra information such as cls_id, score, coordinate

Usage

Method 1: Command Line Usage

python paddleocr_vl_rec.py --image /path/to/your/image.jpg --model_dir /path/to/model

Optional Parameters:

• --model_name: Model name (default: PaddleOCR-VL-0.9B)
• --model_dir: Model directory path (optional)
• --device: Device (e.g., 'cpu', 'gpu:0')
• --prompt_label: Recognition task type (default: 'ocr')
  • Optional values: 'ocr', 'table', 'formula', 'chart'
• --query: Additional custom prompt (optional; if provided, appended after prompt_label)
  • Supports string format in command line
• --max_new_tokens: Maximum number of tokens to generate (default: 4096)
• --return_json: Parse result as JSON format (default: False)

Method 2: Use as a Python Module

from paddleocr_vl_rec import PaddleOCRVLRec

# Initialize recognizer
recognizer = PaddleOCRVLRec(
    model_name="PaddleOCR-VL-0.9B", # Optional, defaults to "PaddleOCR-VL-0.9B"
    model_dir="path/to/your/model",
    device="gpu:0"  # or "cpu"
)

# Recognize text in image (using prompt_label)
result_text = recognizer.predict(
    image="/path/to/your/image.jpg",
    prompt_label="ocr" # Optional, defaults to "ocr"
)

print(result_text)

# Close model
recognizer.close()

Method 3: Using dict or list as query

Note: This method only works with fine-tuned models. The default model cannot correctly recognize and output results in JSON format.

from paddleocr_vl_rec import PaddleOCRVLRec

# Initialize recognizer
recognizer = PaddleOCRVLRec(
    model_dir="path/to/your/model"
)

# Use dict as query (will be converted to JSON string)
# Return JSON format (parsed using json_repair)
result_json = recognizer.predict(
    image="/path/to/your/image.jpg",
    query={"NAME":"", "ITEMS":[]},
    return_json=True
)
# result_json is a dictionary object
print(type(result_json))  # <class 'dict'>
print(result_json)

# Use list as query (will be converted to {"item1":"", "item2":""} format)
result_json = recognizer.predict(
    image="/path/to/your/image.jpg",
    query=["item1", "item2"],
    return_json=True
)
print(result_json)

recognizer.close()

Additional Examples

OCR Text Recognition

python paddleocr_vl_rec.py --image document.jpg --prompt_label ocr

Table Recognition

python paddleocr_vl_rec.py --image table.jpg --prompt_label table

Formula Recognition

python paddleocr_vl_rec.py --image formula.jpg --prompt_label formula

Chart Recognition

python paddleocr_vl_rec.py --image chart.jpg --prompt_label chart

Combining prompt_label and query

# query will be appended after the prompt_label's prompt text
# Final prompt: "OCR:" + "{\"ITEM\":\"\", \"AMOUNT\":\"\"}"
python paddleocr_vl_rec.py --image document.jpg --prompt_label ocr --query "{\"ITEM\":\"\", \"AMOUNT\":\"\"}"