基于深度残差自适应注意力网络的图像超分辨率重建

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

针对现有图像超分辨率重建算法中常见的图像边缘模糊、卷积核尺寸选择单一、重建网络结构冗余等问题,提出一种基于深度残差自适应注意力网络的图像超分辨率重建算法。构建嵌套残差网络结构增加残差网络深度,在保证网络拟合性能的前提下提升网络整体性能。建立自适应注意力模块,使用空洞空间金字塔池化模块融合不同尺度的特征图,获得更多的有效特征,恢复图像纹理细节,同时基于选择性卷积核模块和像素注意力模块的并行结构自适应调整卷积核尺寸,并应用注意力机制提取图像高频特征,最终将提取特征通过重建模块实现超分辨率图像重建。在Set5、Set14、BSD1003个测试数据集上的实验结果表明,与Bicubic、SRCNN、MemNet、DCSR等重建算法相比,该算法的峰值信噪比和结构相似性指标平均提升了0.57 dB和0.0068,具有更高的超分辨率图像重建质量。

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 0.0068 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.