Main Application Goals of GAN:
Generative tasks (generation, reconstruction, super-resolution, style transfer, completion, upsampling, etc.)
The Core Idea of GAN:
Generative model G and discriminator D engage in generations of games
Generator: A generating network that produces images from input
Discriminator: A binary classification network that uses images generated by the generator as negative samples and real images as positive samples
Learn Discriminator D: Given G, generate images through G to create negative samples, and combine them with real images as positive samples to train D
Learn Generator G: Given D, train G to make D’s scoring of images generated by G as close to positive samples as possible
The training process of G and D alternates, and this adversarial process makes the images generated by G increasingly realistic, while D’s ability to