WLSAN: Wavelet-Layer-Spatial Attention Network for Single Image Super Resolution

Abstract

Single Image Super Resolution has been a hot topic in recent years, which has wide application prospects. Some recent works use attention in SR to capture the inter-channel and inter-spatial relationships of feature maps. How- ever, these works cannot focus on the image features at different frequency do- mains. At the same time, the loss function of some SR models will make the model more inclined to reconstruct the smooth part and cannot consider the de- tails in multiple directions of the image. To address these problems, we propose a Wavelet-Layer-Spatial Attention Network (WLSAN), which makes use of both essential global context and local information to improve accuracy and visual quality. Specifically, the proposed WLSAN consists of wavelet attention to model inter-dependencies by decomposing the image signals and non-local attention, including channel-spatial-layers. Meanwhile, we combine L1 with improved Sobel Loss, which enhances the stability during training and constrains different aspects of the HR image generation. Extensive experiments demonstrate that the proposed WLSAN exceeds many classical methods and states of the arts.