labelme-to-datasets-for-OCR-model-training

Convert 'labelme' project's output data to a dataset for OCR model training.

References

1. 'labelme' project: https://github.com/wkentaro/labelme

Procedures

1. Read 'label' and 'bbox' from json file.
2. Crop the image using 'bbox' and save it.
3. Write a label for each cropped image to the label file.

Usage example:

python3 convert.py \
        --input_path ./input \
        --output_path ./output

Input data structure:

/input
├── image00001.png
├── image00001.json
├── image00002.png
├── image00002.json
└── ...

• For the 'images.json' file structure, refer to the 'https://github.com/wkentaro/labelme'

Output data structure:

/output
└── /images
    #   [filename]_[idx].[ext]
    ├── image00001_00001.png
    ├── image00001_00002.png
    ├── image00002_00001.png
    ├── image00002_00002.png
    ├── ...
    └── labels.txt

• Label file structure:

  {filename}\t{label}\n

  image00001_00001.png	abcd
  image00001_00002.png	efgh
  image00002_00001.png	ijkl
  image00002_00002.png	mnop
  ...