hell-to-heaven / handwriting-recognition-for-english Goto Github PK

This repository is handwriting recognition for english.

• python 3.6
• Tensorflow
• keras
• OpenCV

1. prepare fix height train samples for letters, words or sentences.
2. run data_proceess/data_letter_generate.py create single letter or multiple letters samples with 32*120.
3. run data_proceess/data_words_generate.py create english words samples with 64*400.
4. run create_lmdb_dataset.sh can convert image files to lmdb format.

python train.py

You can choose a best model for your work.

1. single letter or multiple letters recognition: simple_densenet.py ====> Densenet+CTC
2. words recognition: densenet_blstm.py ====>Densenet+Bilstm+CTC
3. a new test for cnn: efficientnet_b0.py ====> efficientnet_b0+Bilstm+CTC

python test_model.py

You can choose one method for test.

1. flag=0 # test single or multiple imges
2. flag=1 # test file_lst and compute test precision
3. flag=2 # find hard samples

1. letters: Densenet+CTC 99.8%;
2. words: Densenet+Bilstm+CTC 99.0%.