Second-Order Attention Network for Magnification-Arbitrary Single Image Super-Resolution

Abstract

Recently, owing to the application of deep convolutional neural networks (CNNs), single image super-resolution (SISR) has been developed rapidly. Nevertheless, for low-resolution image magnification, most of the super-resolution methods only take into account the integer scale factor, and there are few methods to consider arbitrary scale factor magnification. Besides, most of the super-resolution methods train separate models for different scale factors, which greatly reduces efficiency. Second-order Attention Network for Single Image Super-Resolution (SAN) is one of the super-resolution models with the best effect, but it cannot achieve magnification-arbitrary super-resolution. To address this issue, we use Meta- Upscale Module as a new upscale module of SAN. This module can realize arbitrary scale magnification by dynamically predicting filter weights through scale factors and position-related vectors. Finally, we propose a super-resolution model based on the SAN feature learning module and Meta-Upscale Module named Meta-SAN. Experiments on Set5, Set14, Urban100, and BSDS100 datasets demonstrate the superiority of our Meta- SAN network.