emmmmmaa/dtr

A network for irregular text recognition.

Deformable Text Recognition

This software implements the Deformable Convolutional Recurrent Neural Network, a combination of of Convolutional Recurrent Neural Network, Deformable Convolutional Networks and Residual Blocks. Some of the codes are from crnn.pytorch and Deformable-ConvNets. For details, please refer to our paper.

Requirements

• Python 3.6
• PyTorch 1.0
• TorchVision
• Numpy
• Six
• Scipy
• LMDB
• Pillow
• warp-ctc-pytorch

Data Preparation

Please convert your own training dataset to LMDB format. The testing images should be in one folder with a txt file, and the txt file contains some lines which contains the filename and the label, formatted as:

img_001.png, "word"

Each line should end with '\n'. You can also use the data provided by us from

• Baidu Netdisk
• Google Drive
• Onedrive

which should be contained in the test_data folder.

Train and Test

To train a new model, simply execute python train.py --lmdb_paths {train_path_list} --cuda. If you need to set other parameters, explore train.py for details.

To test a trained model, you need to explore and execute eval.py.

Citation

@inproceedings{deng2019focus,
  title={Focus-Enhanced Scene Text Recognition with Deformable Convolutions},
  author={Deng, Linjie and Gong, Yanxiang and Lu, Xinchen and Yi, Xin and Ma, Zheng and Xie, Mei},
  booktitle={2019 IEEE 5th International Conference on Computer and Communications (ICCC)},
  pages={1685--1689},
  year={2019},
  organization={IEEE}
}