Please find the orginal implementation at CRNN_Tensorflow. I have made some changes to support all ascii characters and output a confidence score for each recognized word. The model uses a CRNN architecure which includes a CNN, LSTM and CTC loss.The whole project is wrapped up end to end as a web seloution.
I have trained the model on my collection of data from pdfs. There are othere available datasets online like Synth 90k. I have also generated a lot of syntetic data using Text-genrator.
- Collect as much data as possible, put them in dataset/Train, dataset/Test directories. Include a text file sample.txt in which each row contains an image name and its label
- Run tools/write_text_features.py to generate tfrecords for training, validation and testing. All the images will be resized to 100*32
python tools/write_text_features.py
- Run training script
python tools/train_shadownet.py
- Orginal CRNN paper:http://arxiv.org/abs/1507.05717.
- CRNN implementation: https://github.com/bgshih/crnn
- Data generation: https://github.com/Belval/TextRecognitionDataGenerator