Convolutional Recurrent Neural Network

This software implements the Convolutional Recurrent Neural Network (CRNN), a combination of CNN, RNN and CTC loss for image-based sequence recognition tasks, such as scene text recognition and OCR. For details, please refer to our paper http://arxiv.org/abs/1507.05717.

Build

The software has only been tested on Ubuntu 14.04 (x64). CUDA-enabled GPUs are required. To build the project, first install the latest versions of Torch7, fblualib and LMDB. Please follow their installation instructions respectively. On Ubuntu, lmdb can be installed by apt-get install liblmdb-dev.

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

Run demo

A demo program can be found in src/demo.lua. Before running the demo, download a pretrained model from here. Put the downloaded model file crnn_demo_model.t7 into directory model/crnn_demo/. Then launch the demo by:

th demo.lua

The demo reads an example image and recognizes its text content.

Example image:

Expected output:

Loading model...
Model loaded from ../model/crnn_demo/model.t7
Recognized text: available (raw: a-----v--a-i-l-a-bb-l-e---)

Use pretrained model

The pretrained model can be used for lexicon-free and lexicon-based recognition tasks. Refer to the functions recognizeImageLexiconFree and recognizeImageWithLexicion in file utilities.lua for details.

Train a new model

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

Create a new LMDB dataset. A python program is provided in tool/create_dataset.py. Refer to the function createDataset for details.
Create model directory under model/. For example, model/foo_model. Then create configuraton file config.lua under the model directory. You can copy model/crnn_demo/config.lua and do modifications.
Go to src/ and execute th main_train.lua ../models/foo_model/. Model snapshots and logging file will be saved into the model directory.

Citation

Please cite the following paper if you are using the code/model in your research paper.

@article{ShiBY15,
  author    = {Baoguang Shi and
               Xiang Bai and
               Cong Yao},
  title     = {An End-to-End Trainable Neural Network for Image-based Sequence Recognition
               and Its Application to Scene Text Recognition},
  journal   = {CoRR},
  volume    = {abs/1507.05717},
  year      = {2015}
}

Acknowledgements

The authors would like to thank the developers of Torch7, TH++, lmdb-lua-ffi and char-rnn.

Please let me know if you encounter any issues.