Class-Consistency CycleGAN restrict generated data by given labels. Derived from original CycleGAN, we insert two classification Networks with shared weights as the two constrain branches.
In this repo, I generate license plate and restricted the generated license plate characters by inserting a crnn network as the classifcation Network and training the branch with ctc-loss.
The license plate numbers in proposed CC-CycleGAN are clear and recognizable.This is highly based on the subfolder "cyclegan" in https://github.com/eriklindernoren/Keras-GAN repo.