This is the implementation of the paper "Gated Recurrent Convolution Neural Network for OCR"
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.

Gated Recurrent Convolution Neural Network for OCR

This project is an implementation of the GRCNN for OCR. For details, please refer to the paper:


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:

It is more convenient if you use nivdia-docker image (@rremani supplied) :

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


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/ 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.


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 so that the shape of receptive field is similar to the character.


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

Other applications

GRCNN for face recognition. We build a GRCNN and use a modified combined angular margin softmax loss for face recognition. GRCNN can reach state-of-the-art results on lfw (99.82 %)

The code of our journal version paper which applies GRCNN to object recognition can be found at:

The pretrained model on ImageNet will be released.


Caffe implementation: Since the CRNN project is a little bit out-of-date, it might cannot compatible with Ubuntu 16.04. We will further implement a caffe version of GRCNN for OCR.