This repository is handwriting recognition for english.
- python 3.6
- Tensorflow
- keras
- OpenCV
- prepare fix height train samples for letters, words or sentences.
- run data_proceess/data_letter_generate.py create single letter or multiple letters samples with 32*120.
- run data_proceess/data_words_generate.py create english words samples with 64*400.
- run create_lmdb_dataset.sh can convert image files to lmdb format.
python train.py
You can choose a best model for your work.
- single letter or multiple letters recognition: simple_densenet.py ====> Densenet+CTC
- words recognition: densenet_blstm.py ====>Densenet+Bilstm+CTC
- a new test for cnn: efficientnet_b0.py ====> efficientnet_b0+Bilstm+CTC
python test_model.py
You can choose one method for test.
- flag=0 # test single or multiple imges
- flag=1 # test file_lst and compute test precision
- flag=2 # find hard samples
- letters: Densenet+CTC 99.8%;
- words: Densenet+Bilstm+CTC 99.0%.