CRNN_Chinese_Characters_Rec

Characters Recognition

A Chinese characters recognition repository based on convolutional recurrent networks. (Below please scan the QR code to join the wechat group.)

Performance

Recognize characters in pictures

Dev Environments

1. WIN 10 or Ubuntu 16.04
2. PyTorch 1.2.0 (may fix ctc loss) with cuda 10.0 🔥
3. yaml
4. easydict
5. tensorboardX

Data

Synthetic Chinese String Dataset

1. Download the dataset

2. Edit lib/config/360CC_config.yaml DATA:ROOT to you image path

   DATASET:
     ROOT: 'to/your/images/path'
   

3. Download the labels (password: eaqb)

4. Put char_std_5990.txt in lib/dataset/txt/

5. And put train.txt and test.txt in lib/dataset/txt/

   eg. test.txt

   20456343_4045240981.jpg 89 201 241 178 19 94 19 22 26 656
   20457281_3395886438.jpg 120 1061 2 376 78 249 272 272 120 1061
   ...
   

   Or your own data

6. Edit lib/config/OWN_config.yaml DATA:ROOT to you image path

   DATASET:
     ROOT: 'to/your/images/path'
   

7. And put your train_own.txt and test_own.txt in lib/dataset/txt/

   eg. test_own.txt

   20456343_4045240981.jpg 你好啊！祖国！
   20457281_3395886438.jpg 晚安啊！世界！
   ...
   

   note: fixed-length training is supported. yet you can modify dataloader to support random length training.

Train

   [run] python train.py --cfg lib/config/360CC_config.yaml
or [run] python train.py --cfg lib/config/OWN_config.yaml

#### loss curve

angular2html [run] cd output/360CC/crnn/xxxx-xx-xx-xx-xx/ [run] tensorboard --logdir log

#### loss overview(first epoch)
<center/>
<img src='images/train_loss.png' title='loss1' style='max-width:800px'></img>
</center>
<p>
<img src='images/tb_loss.png' title='loss1' style='max-width:600px'></img>
</p>

## Demo

angular2html [run] python demo.py --image_path images/test.png --checkpoint output/checkpoints/mixed_second_finetune_acc_97P7.pth ```

References

• https://github.com/meijieru/crnn.pytorch
• https://github.com/HRNet