Sliding Convolution CTC for Scene Text Recognition

Implementation of 'Scene Text Recognition with Sliding Convolutional Character Models'(pdf)

Model

Sliding windows + CNN + CTC

Dependency

While this implement might work for many cases, it is only tested for environment below:

python == 3.7.0
torch == 0.4.1
tqdm
numpy

warp-ctc(for pytorch 0.4)

CUDA 9.0.1
CUDNN 7.0.5

Install warp-ctc

Follow this instruction

Note:Version of warp-ctc should be corresponding with pytorch. Related issue

Usage

Download IIIT5K dataset and release files to dataset folder.

Preprocess IIIT5K dataset

python3 prepare_IIIT5K_dataset.py

Train model:

python3 main.py --cuda=True --mode=train

Resume training:

python3 main.py --cuda=True --wram-up=True --mode=train

Test model:

python3 main.py --cuda=True --mode=test

Note: model.bin file is a pre-trained model which could achieve about 53% accuracy. (Due to the small training dataset)

Citation

If you find this work is useful in your research, please consider citing:

@article{yin2017scene,
  title={Scene text recognition with sliding convolutional character models},
  author={Yin, Fei and Wu, Yi-Chao and Zhang, Xu-Yao and Liu, Cheng-Lin},
  journal={arXiv preprint arXiv:1709.01727},
  year={2017}
}