Cooperative Adversarial Network for Accurate Super Resolution

Abstract

Convolutional neural networks for super-resolution (SR) have recently improved global pixel similarity. To estimate visually pleasing SR images, general SR methods focus on optimizing a network by minimizing the feature loss provided by a pre-trained classification network. These estimated SR images are usually enriched with high-frequency texture details, but accompanied with large global pixel error. In this paper, we design a cooperative adversarial network (CAN) for SR, which consists of two sub-networks: chaser and enhancer. The chaser estimates visually pleasing SR images by optimizing feature loss provided by the enhancer, and the enhancer promotes the SR images of the chaser with less global pixel error. In addition, the enhancer fuses multiple residual features from different layers to reserve different level details for the SR images. In order to merge the two sub-networks into a unified network, we define an equilibrium ratio of the pixel-wise loss of chaser to enhancer for controlling alternate process of three training units. Compared to existing multi-network models, our CAN shows competitive performance with less parameters.