Adaptive diagonal total-variation generative adversarial network for super-resolution imaging

Abstract

To address problems that the loss function does not correlate well with perceptual vision in super-resolution methods based on the convolutional neural network(CNN), a novel model called the ADTV-SRGAN is designed based on the adaptive diagonal total-variation generative adversarial network. Combined with global perception and the local structure adaptive method, spatial loss based on the diagonal variation model is proposed to make the loss function can be adjusted according to the spatial features. Pixel loss and characteristic loss are in combination with the spatial loss for the fusing optimization of the total loss function such that high-frequency details of the images are maintained to improve their quality. The results of experiment show that the proposed method can obtain competitive results in objective evaluations. In subjective assessment, images reconstructed by it are clear, delicate, and natural, and it preserved edge- and texture-related details.