This software implements the Convolutional Recurrent Neural Network (CRNN), a combination of CNN and Sequence to sequence model with attention for image-based sequence recognition tasks, such as scene text recognition and OCR.

This network architecture is implemented from Robust Scene Text Recognition with Automatic Rectification

The crnn+ctc ocr can be found here bai-shang/crnn_ctc_ocr.Tensorflow

All dependencies should be installed are as follow:

• Python3.5
• PyTorch
• opencv-python
• numpy
• Pillow

Required packages can be installed with

pip3 install -r requirements.txt

This software only tested in Python3.5!

Asume your current work directory is "crnn_seq2seq_ocr.Pytorch"：

#cd crnn_seq2seq_ocr.Pytorch
python3 inference.py --img_path ./data/test_img/20439171_260546633.jpg \
    --encoder model/pretrained/encoder.pth --decoder model/pretrained/decoder.pth

Result is:

predict_string: 于采取的是“激变”的 => predict_probility: 0.860094428062439

• Download Synthetic Chinese String Dataset.

• Create train_list.txt and test_list.txt as follow format

path/to/your/image/50843500_2726670787.jpg 情笼罩在他们满是沧桑
path/to/your/image/57724421_3902051606.jpg 心态的松弛决定了比赛
path/to/your/image/52041437_3766953320.jpg 虾的鲜美自是不可待言

You can use the "data/convert_text_list.py" script to create the two lists or finish it by yourself.

cd data
python3 convert_text_list.py SyntheticChineseStringDataset/train.txt > train_list.txt
python3 convert_text_list.py SyntheticChineseStringDataset/test.txt > test_list.txt

• Start training

#cd crnn_seq2seq_ocr.Pytorch
python3 --train_list train_list.txt --eval_list test_list.txt --model ./model/crnn/ 

Then the training messages are printed to terminal like

• caffe_ocr

• PyTorch Tutorials > Translation with a Sequence to Sequence Network and Attention