Need help with crnn-mxnet-chinese-text-recognition?
Click the “chat” button below for chat support from the developer who created it, or find similar developers for support.

About the developer

diaomin
151 Stars 293 Forks 16 Commits 13 Opened issues

Description

An implementation of CRNN (CNN+LSTM+warpCTC) on MxNet for chinese text recognition

Services available

!
?

Need anything else?

Contributors list

crnn-mxnet-chinese-text-recognition

This is an implementation of CRNN (CNN+LSTM+CTC) for chinese text recognition.

Building MXNet with warp-ctc

  1. In order to use
    mxnet.symbol.WarpCTC
    layer, you need to first build Baidu's warp-ctc library from source
  2. Then build MXNet from source with warp-ctc config flags enabled.

Data Preparation

  1. Download the Synthetic Chinese Dataset(contributed by https://github.com/senlinuc/caffe_ocr and many thanks)

A glance of the dataset: * almost 3.6 million synthetic chinese text images. * 5,990 different categories in total. * each image has a length of 10 characters.

  1. Create train.txt and text.txt with the format like this:
           image_name1 label1_1 label1_2 label1_3...
           image_name2 label2_1 label2_2 label2_3...
    
    Optional: downoad the two files here

Training

  1. Modify the path of images and txt files in train.py
  2. Run
    $ python train.py 2>&1 | tee log.txt
    
  3. After almost 19 epoches, you can get 99.0502% validation accuracy.
    2018-04-01 03:35:35,136 Epoch[18] Batch [25450] Speed: 53.10 samples/sec    accuracy=0.988125
    2018-04-01 03:37:37,482 Epoch[18] Batch [25500] Speed: 52.31 samples/sec    accuracy=0.986719
    2018-04-01 03:39:38,613 Epoch[18] Batch [25550] Speed: 52.84 samples/sec    accuracy=0.989531
    2018-04-01 03:41:40,470 Epoch[18] Batch [25600] Speed: 52.52 samples/sec    accuracy=0.987969
    2018-04-01 03:42:27,544 Epoch[18] Train-accuracy=0.988672
    2018-04-01 03:42:27,544 Epoch[18] Time cost=80796.510
    2018-04-01 03:42:27,610 Saved checkpoint to "./check_points/model-0019.params"
    2018-04-01 05:34:43,096 Epoch[18] Validation-accuracy=0.990502
    
    Hare is a pre-trained model you can download directly.

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.