SRGAN with Total Variation Loss in Face Super-Resolution

Facial image super-resolution is a crucial preprocessing for facial image analysis, face recognition, and image-based 3D face reconstruction. Convolutional neural networks were earlier used to produce high-resolution images that train quicker and shown excellent performance by learning mapping relation using pairs of low-resolution and high-resolution images. However, in some cases, they are incapable of recovering finer details and often generate blurry images. In this paper, we evaluate a method of applying Generative adversarial networks in generating realistic super-resolution images from low-resolution ones by using three typical losses for super-resolution: Content Loss, Adversarial Loss, Perceptual Loss, and proposed to use Total Variation Loss. We try different pre-trained famous Convolutional neural networks models (VGG19, FaceNet, and EfficientNet) in Perceptual Loss to have a general view with different backbones. Our network gains 32.67 of Peak signal-to-noise ratio (PSNR) and 0.89 of Structural similarity index (SSIM) in 100 random samples from the Flickr-Faces-HQ dataset.