Agora, Inc.
Effective structure keeping for generative adversarial networks for single image super resolution

Abstract:

A method of training a generator G of a Generative Adversarial Network (GAN) includes generating a real contextual data set {x.sub.1, . . . , x.sub.N} for a high resolution image Y; generating a generated contextual data set {g.sub.1, . . . , g.sub.N} for a generated high resolution image G(Z); calculating a perceptual loss L.sub.pcept value using the real contextual data set {x.sub.1, . . . , x.sub.N} and the generated contextual data set {g.sub.1, . . . , g.sub.N}; and training the generator G using the perceptual loss L.sub.pcept value. The generated high resolution image G(Z) is generated by the generator G of the GAN in response to receiving an input Z, where the input Z is a random sample that corresponds to the high resolution image Y.