Image Super-Resolution Reconstruction Based on Depth Residual Adaptive Attention Network

Abstract

To solve problems typically encountered in existing image Super-Resolution（SR） reconstruction algorithms，such as blurred image edges，single selection of convolution kernel size，and redundant reconstruction network structure，this paper proposes an image SR reconstruction algorithm based on a Deep Residual Adaptive Attention Network（DRAAN）.Applying a Residual in Residual（RIR） network structure increases the depth of the residual network and improves the overall performance by ensuring the fitting performance of the network.In the DRAAN，an Adaptive Attention（AA） module is established，and an Atrous Spatial Pyramid Pooling（ASPP） module is used to fuse feature maps of different scales，obtain effective features，and restore image texture details.Additionally，based on the parallel structure of the Select Kernel（SK） and Pixel Attention（PA） modules，the size of the convolution kernel is adaptively adjusted，and the attention mechanism is applied to efficiently extract the high-frequency features of images.Finally，to achieve SR image reconstruction，the extracted features are reconstructed via module reconstruction.Test and simulation results on three benchmark datasets，Set5，Set14，and BSD100，show that compared with Bicubic，deep Convolutional Network for image Super-Resolution（SRCNN），persistent Memory Network（MemNet） for image restoration，Dilated Convolutions for single-image Super-Resolution（DCSR），and other reconstruction algorithms，the proposed algorithm yields 0.57 dB and 0.006 8 higher values of Peak Signal to Noise Ratio（PSNR） and Structural SIMilarity（SSIM），respectively，on average as well as higher a SR image reconstruction quality.