for basic discriminator of GANs
Discriminator for sensing the generated composite image and the generated difference of ground-truth, and aimed to distinguish the fake or real;
Meanwhile, the output of the discriminator is passed through a series of scalar values after conv obtained, so that generally the activation value between 0 and 1;
However, such a result there are some problems:
1. The results clearly output a weighted value of the whole picture, not embodying features of the partial images, a high precision is difficult for an image migration task.
for Patch-based discriminator of GANs
PatchGAN idea is the final output is not a scalar value, but a $ N * N $ matrix $ X $, in fact,
There solution is to cut the image into a plurality of patches, respectively discriminator identifying differences, but the presence of such a large consumption calculation.