Face Image Super-resolution Based On Relative Average Generative Adversarial Networks

Abstract

Face image plays an important role in visual perception and various computer vision tasks. However, due to the influence factors of equipment and environment, the image often has the problem of low resolution. In order to relieve this problem, this paper proposes a face image super-resolution reconstruction algorithm based on Relative average Generative Adversarial Networks. Different from the standard discriminator D in Generative Adversarial Networks(GAN), which estimates the probability that one input image is real and natural, a relativistic average discriminator tries to predict on average the probability that a real image is relatively more realistic than a fake one. At the same time, the loss function used to measure the spatial similarity of image pixels is combined with the perceptual loss function used to measure the similarity of image feature spaces, so that the network pays attention to the reconstruction of image pixel information while taking into account the feature information of the image. Experimental results demonstrate the effectiveness of the proposed method in multi-scale face image super-resolution, and the evaluation indicators (PSNR and SSIM) tested on common test sets are better than those of the contrast algorithm, it also has better visual perception and more detailed information.