number detection and recognition based on AdvancedEast and CRNN
Detection and Crop:
Recognition:
- 22046298859.jpg : 22046298859
- 97785067838.jpg : 97785067838
- 84999825604.jpg : 84999825604
- 99851544924.jpg : 99851544924
- 28510715459.jpg : 28510715459
- 12233418739.jpg : 12233418739
- 41679405336.jpg : 41679405336
- 37774346979.jpg : 37774346979
limitations: When the two models are test on their respective validation sets , they can reach an acc of about 0.9. However, the number of the training data for recognizer I generated is horizontal, and the number in the crop image after the detection result introduces the rotation and other factors, resulting in poor results when used in combination.
- prepare training data, data format refer to ICPR
- modify params in cfg.py
- run python preprocess.py to resize image and generator .npy training files
- run python label.py
- run python train.py, train the network
- modify your images' dir in predict.py, and run python predict.py, then we will get three outputs: bounding box on origin images, the cropped image, and coordinates(txt file).
more details please refer to AdvancedEast
- prepare training data, data format refer to MJSynth data
- modify params in cfg.py
- modify input_shape=(None, 50,7,512) in train.py line 55, the input_shape is refer to your bn_shape = bn4.get_shape() in network.py
- run python train.py
- modify your images' dir in predict.py, then run python predict.py