基于残差注意力网络的图像超分辨率重建方法

Single-image Super-resolution Reconstruction Algorithm Based on Residual Attention Network

针对自然图像超分辨率重建中纹理细节模糊问题,提出一种基于残差注意力网络的重建方法。在特征提取网络层中添加通道注意力块,自适应确定局部跨通道的多维范围,提高特征表达能力,使网络能学习到通道中更多的高频信息;添加残差学习模块,通过捕捉易丢失的高频残差信息得到更多的语义表达信息;最终通过上采样输出高分辨率重建图像。实验结果表明,该方法在自然图像超分辨率重建上有良好的性能,能重建出更清晰的图像。

A reconstruction approach based on residual attention network is developed to address the problem of texture detail blur in super-resolution reconstruction of natural images.In the feature extraction network layer,add channel attention blocks to adaptively determine the multi-dimensional range of local cross channels,increase feature expression,and allow the network to learn more high-frequency information in the channel;By capturing the easily lost high-frequency residual information,the residual learning module can provide more semantic expression information.Finally,up sampling is used to output the high-resolution reconstructed image.Experiments reveal that this method performs well in super-resolution natural image reconstruction and can recreate crisper images.