Skip to content

Latest commit

 

History

History
73 lines (41 loc) · 2.6 KB

README.md

File metadata and controls

73 lines (41 loc) · 2.6 KB
Raw

Gated Recurrent Convolution Neural Network for OCR

This project is an implementation of the GRCNN for OCR. For details, please refer to the paper: https://papers.nips.cc/paper/6637-gated-recurrent-convolution-neural-network-for-ocr.pdf

Update

The journal version of GRCNN has been accepted by T-PAMI 2021, and the code is available at:

https://github.com/Jianf-Wang/GRCNN

Build

The GRCNN is built upon the CRNN. The requirements are:

  1. Ubuntu 14.04
  2. CUDA 7.5
  3. CUDNN 5

For the convenience of compiling, we provide the dependencies from here: https://pan.baidu.com/s/1c21zl1e#list/path=%2F

It is more convenient if you use nivdia-docker image (@rremani supplied) : https://hub.docker.com/r/rremani/cuda_crnn_torch/

After installing the dependencies, go to src/ and execute build_cpp.sh to build the C++ code. If successful, a file named libcrnn.so should be produced in the src/ directory.

Inference

We provide the pretrained model from here. Put the downloaded model file into directory model/GRCL/. Moreover, we provide the IC03 dataset in the "./data/IC03" directory. You need to change the directories listed in the "test.txt". The "test_label.txt" is the ground truth of each image. The "lexicon_50.txt" is the lexicon of IC03.

"src/evaluation.lua": Lexicon-free evaluation

"src/evaluation_lex.lua" Lexicon-based evaluation

The evaluation code will output the recognition accuracy.

Train a new model

Follow the following steps to train a new model on your own dataset.

  1. Create a new LMDB dataset.src/create_own_dataset.py(need to pip install lmdb first).
  2. You can modify the configuration in model/GRCL/GRCL_LSTM_pretrain.lua
  3. Go to src/ and execute th main_train.lua ../model/GRCL/ ../model/saved_model. Model snapshots will be saved into ../model/saved_model.

Visualization

We visualize the RCNN , DenseNet and GRCNN to verify the dynamic receptive fields in GRCNN for OCR. There are clearly gaps among different characters, and for each character, the unrelated parts do not provide strong signal.

Citation

@inproceedings{jianfeng2017deep,
 author    = {Wang, Jianfeng and Hu, Xiaolin},
 title     = {Gated Recurrent Convolution Neural Network for OCR},
 booktitle = {Advances in Neural Information Processing Systems},
 year      = {2017}
}