Blurred image restoration conditional generative adversarial network based on hybrid channel and spatial attention

Abstract

In this paper, we propose a conditional generative adversarial network (CGAN) for restoring blurred image. The design of the generator derives from classic U-net network, but to improve its expression ability, we first modify the U-net by replacing some deep layers with stacked residual modules. Furthermore, we combine the channel and spatial attention modules and embed them into the generator to force it paying more attention to important channels and blurred local space. For loss function design, we comprehensively incorporate the pixel loss, perception loss and adversarial loss to enhance the performance of the proposed CGAN. Finally, the GoPro dataset is used for training and evaluating the effectiveness of the network. The results show that the proposed CGAN can achieve restored image of very high quality which is comparable with some state of the art methods.