PPOCRLabel

PPOCRLabelv3 is a semi-automatic graphic annotation tool designed for the OCR field. It features built-in PP-OCR models to automatically detect and re-recognize data. The tool is written in Python3 and PyQT5 and supports several annotation modes:

• Regular text annotation: Rectangular box labeling.
• Table annotation: Specialized mode for table structures.
• Irregular text annotation: Multi-point labeling for non-rectangular text.
• Key information annotation (KIE): For detection, recognition, and keyword extraction.

Annotations produced can be used directly for training PP-OCR detection and recognition models.

Unreadable Text (### vs *)

For the PaddleOCR detection training pipeline, use the transcription string ### or * to identify regions that should be ignored. For maximum compatibility across both Detection and Recognition fine-tuning stages, it is recommended to always use ### for unreadable text.

The difficult field

The exported Label.txt contains a difficult field. This is a legacy field from labelImg. In PPOCRLabel, this field is hardcoded to False for all exported labels to ensure backward compatibility with older data pipelines and prevent unintended data exclusion.

By default, PPOCRLabel uses ultra-lightweight Chinese and English models. You can switch languages (French, German, Korean, Japanese) via the menu bar: PaddleOCR -> Choose OCR Model.

To use your own custom inference models, you can pass the model directories and names as command-line arguments when starting the application.

python PPOCRLabel.py --det_model_dir {your_det_model_dir} --det_model_name {your_det_model_name} --rec_model_dir {your_rec_model_dir} --rec_model_name {your_rec_model_name}

If you have modified the PPOCRLabel source code (e.g., to use custom built-in models) or are performing secondary development, run the application directly using the Python script instead of the installed whl package.

cd ./PPOCRLabel
python PPOCRLabel.py --lang ch

Use the gen_ocr_train_val_test.py script to split your labeled dataset. This script expects a specific directory structure containing crop_img, Label.txt, and rec_gt.txt.

cd ./PPOCRLabel
python gen_ocr_train_val_test.py --trainValTestRatio 6:2:2 --datasetRootPath ../train_data

The easiest way to use PPOCRLabel is by installing the pre-built wheel package. This method automatically handles dependencies like paddleocr.

Windows

pip install PPOCRLabel

Note: If you encounter [winError 126] 找不到指定模块 related to shapely, manually download and install the shapely wheel from Gohlke's Python libs.

Ubuntu Linux

pip3 install PPOCRLabel
pip3 install trash-cli
export QT_QPA_PALTFORM = wayland

MacOS

pip3 install PPOCRLabel
pip3 install opencv-contrib-python-headless==4.2.0.32

# Windows example
pip install PPOCRLabel

To package PPOCRLabel into a standalone executable, follow these steps to regenerate resources and run the PyInstaller build command. This is useful for distributing the tool without requiring a Python environment.

cd ./PPOCRLabel
# 1. Install PyInstaller
pip install pyinstaller

# 2. Regenerate resources
pyrcc5 -o libs/resources.py resources.qrc

# 3. Package the executable
pyinstaller -c PPOCRLabel.py --collect-all paddleocr --collect-all pyclipper --collect-all imghdr --collect-all skimage --collect-all imgaug --collect-all scipy.io --collect-all lmdb --collect-all paddle --hidden-import=pyqt5 -p ./libs -p ./ -p ./data -p ./resources -F

# 4. Run the executable (Windows example)
PPOCRLabel.exe --lang ch

PPOCRLabel exports data in two ways:

Manual Export

Use the menu File -> Export Label to manually export detection results. The results are stored in Label.txt within the opened picture folder.

Automatic Export

If File -> Auto export Label Mode is selected, the program exports automatically. Manual checks are stored in Label.txt.

Recognition Training Data Export

To export data specifically for recognition model training:

1. Click File -> Export Recognition Results in the menu bar.
2. The recognition training data (cropped images) will be saved in the crop_img folder.
3. The recognition labels will be saved in rec_gt.txt.

If you have modified the source code (e.g., to specify new built-in models), run the application directly using the Python script instead of the whl package.

cd ./PPOCRLabel

# Normal mode
python PPOCRLabel.py

# KIE mode
python PPOCRLabel.py --kie True

To package the application into a standalone executable using Pyinstaller, follow these steps:

1. Install Pyinstaller.
2. Regenerate the resources.
3. Run the Pyinstaller command with necessary collection flags for dependencies.

cd ./PPOCRLabel
pip install pyinstaller

# Regenerate Resources
pyrcc5 -o libs/resources.py resources.qrc

# Packaging executable programs
pyinstaller -c PPOCRLabel.py --collect-all paddleocr --collect-all pyclipper --collect-all imghdr --collect-all skimage --collect-all imgaug --collect-all scipy.io --collect-all lmdb --collect-all paddle --hidden-import=pyqt5  -p ./libs -p ./ -p ./data -p ./resources -F

After building, you can run the executable (e.g., PPOCRLabel.exe --lang ch on Windows) from the dist directory.