Self-distillation Regularized Connectionist Temporal Classification Loss for Text Recognition: A Simple Yet Effective Approach

Paper: https://arxiv.org/abs/2308.08806

This an implementation based on PyTorch only. In the paper, we implemented dctc using cuda based on PaddleOCR. As you can see below, DCTC can achieve better performance and converge much faster than ctc.

Experiments

In this repo, we only train the CRNN on the Document split of Chinese Benchmark Dataset and use its codebase. We use one Nvidia A6000 with batch size 512 and 125 epochs. All the logs and ckpts are here.

Model	Document
CRNN (CTC)	96.474%
CRNN (DCTC)	97.516%

Use DCTC in your project

import dctc

criterion = dctc.DCTC(use_il=False, alpha=0.01)

...
...

targets_dict = {
            'targets': text,
            'target_lengths': length
        }
cost = criterion(logits=preds,
                 targets_dict=targets_dict)

Training instructions

1. perpare Chinese Benchmark Dataset
2. cd benchmarking-chinese-text-recognition\model\CRNN
3. modify dataset path in the train.sh

If you find it uesful, please cite our paper.

@misc{zhang2023selfdistillation,
      title={Self-distillation Regularized Connectionist Temporal Classification Loss for Text Recognition: A Simple Yet Effective Approach}, 
      author={Ziyin Zhang and Ning Lu and Minghui Liao and Yongshuai Huang and Cheng Li and Min Wang and Wei Peng},
      year={2023},
      eprint={2308.08806},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}