cnn + lstm/gru + ctc (CRNN) for image text recognition
test results after training 300 steps:
To run this repo:
1, python data_detect_generator.py 0 # to generate validation images for detection
2, python data_detect_generator.py 1 # to generate training images for detection
3, python data_rects_extractor.py 0 # to generate validation data-rects for recognition
4, python data_rects_extractor.py 1 # to generate training data-rects for recognition
5, python script_recog.py # to train and validate
By 1 and 2, images with texts will be generated with images in images_base as background. The images will be saved in the newly-maked folders: ./data_train and ./data_valid.
By 3 and 4, text-rects will be extracted from images generated by 1 and 2. The rect images and texts will be saved in the newly-maked folders: ./data_rects_train and ./data_rects_valid.
By 5, the model will be trained and validated. The ckpt files will be stored in the newly-maked directory, ./model_recog.
The model is mainly based on the method described in the article:
An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application to Scene Text Recognition
Baoguang Shi, Xiang Bai, Cong Yao
We thank Jerod Weinman for making his code available: https://github.com/weinman/cnn_lstm_ctc_ocr