general character recogonizer
- Tensorflow
- PIL
This project's destination is to recogonize any character sequence in one image.
Util reaching the final goal, it may need a number of steps. Espeacially, the final result may not satisfy us.
Charactor recogonition task is very hard in uncertain context, the training set is not same with the test set.
In this project, I will adopt a new policy to do this thing.
Firstly, I will train the network in some steps seperately with a generated training set which will be not related with the captcha world.
Secondly, I will try to introduce attention mechinism into the Neural Network.
Thirdly, I will employ Dynamic Conventional Neural Network(DCNN) and give the network time to focus, think and infer.
Fourthly, I will visualize all the middle layers so that the NN is under my control.
The Generator will generates all the training sets, the Trainer will train every layer for DCNN, the Runner will run the DCNN to do the captha recognition tasks.
- util.py - common utils for images and text
- line.py - line drawer
- img.py - Image utils
- generator.py - The Generator generates all kinds of characters
- trainer.py - The Trainer
- runner.py - The Runner
- nn.net - The trained and saved network parameters
- sparse_auto_encode.py - Implementation of Sparse Auto Encoder. It will be used to train some layers under unsupervised learning, which will construct the basic layers of the whole Neural Network.
- tester.py - test basic functions and features