mxnet-cnn-lstm-ctc-ocr

This repo contains code written by MXNet for ocr tasks, which uses an cnn-lstm-ctc architecture to do text recognition.

In addition buctketing module is used in the code to handle variable length of input images.

network

The network in this preject is based on An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application to Scene Text Recognition. B. Shi， X. Bai， C. Yao .TPAMI

paper https://arxiv.org/abs/1507.05717 original code written by torch https://github.com/bgshih/crnn

the main difference is that I use resnet as the cnn part in the architecture.

prerequisites

1. you should follow steps in official mxnet example/warp-ctc https://github.com/dmlc/mxnet/tree/master/example/warpctc    to make sure you install warp-ctc correctly and recompile mxnet with warp-ctc plug-in
2. download ICDAR2013 cropped word dataset http://rrc.cvc.uab.es/?ch=2&com=downloads and put it in the right fold    which should be consistant with the path in text_deep_ocr_bucketing.py

  path='crop_icdar2013_train.lst'
  path_test='crop_icdar2013_val.lst'
  data_root='/cache/icdar2013_word'
  test_root='/cache/icdar2013_word'

train the model

  make_list.py is to generate .lst file for recording image paths and corresponding labels and idx2char.json,char2idx.json for recording index and char's correspondence , run this in your terminal:

mkdir model
python text_deep_ocr_bucketing.py

results

if you use the default setting your should reach obout 40% accuracy in val set. To improve performance in future work:

1. use synthetic 90k to pretrain the model and finetune on ICDAR2013
2. use data argumentation