Brief introduction of cycle GAN

cycle GAN

This GAN is suitable for the conversion of unpaired images to images. The style transfer
p2pGAN must be trained on paired data. In real life, it is more difficult to find the
principle: get the features of one data set and transform it into the features of another data set. Actually The goal is to learn the mapping from X to Y. We set this mapping as F. F corresponds to the generator of GAN. F can transform the picture x in X into the picture F(x) in Y. For the generated picture, we Need to use a discriminator to judge whether it is a real rain image.
Then there will be a problem: if we use the traditional generation of confrontation network to separate LOSS, the result we get will become a picture instead of the desired image, and the loss will be nullified. That is: mapping F can completely map all x to the same picture in Y space, and all input images become an image in Y, which also meets the standard of D and invalidates the loss.
This author proposed In order to improve the cycle consistency, the

paper uses the generated image to reconstruct and compare it with an image in A. If the same, we then input it to the discriminator
cycle GAN for color and texture conversion, but the conversion effect for image graphics is not good.

Data set download: https://blog.csdn.net/zhou_438/article/details/105290100