多尺度混合注意力网络的图像超分辨率重建

Multi-scale hybrid attention network for image super-resolution reconstruction

针对现有超分辨率(super-resolution)重建算法重建出的图像存在高频细节丢失、结构化失真的问题,结合多尺度混合注意力网络,给出一种新的重建算法。首先,设计了一种多尺度残差模块(multi-scale residual module,MRM),提取不同尺度信息的特征并进行融合来获取包含更多信息的浅层特征;其次,采用残差混合注意力模块(residual hybrid attention module,RHAM),依次沿着通道和空间2个不同的维度增强网络特征提取能力,进行自适应的特征优化,提高高频特征的复用;最后,通过重建模块对提取的特征进行增强,获取相应的高分辨率图像。在基准数据集上进行测试,实验结果表明:文中提出的算法相较主流图像SR算法,在放大尺度为2、3、4倍时峰值信噪比(peak signal to noise ratio,PSNR)平均提高了0.104、0.224、0.146 dB,结构相似性(structural similarity index measure,SSIM)平均提高了0.0349、0.0276、0.0181。该算法能更有效地利用原始图像信息,重建出的图像在边缘和纹理细节等方面有一定的提高。

Aiming at the problems of high-frequency detail loss and structural distortion in images reconstructed by existing super-resolution(SR)reconstruction algorithms,a new reconstruction algorithm is proposed by combining multi-scale hybrid attention network.Firstly,a multi-scale residual module(MRM)was designed to extract features of different scale information and fuse them to obtain shallow features containing more information.Secondly,the residual hybrid attention module(RHAM)was used to enhance the network feature extraction ability along two different dimensions of channel and space,and adaptive feature optimization was carried out to improve the reuse of high-frequency features.Finally,the extracted features were enhanced through the reconstruction module to obtain corresponding high-resolution images.Tested on a benchmark data set,the experimental resulTSshow that the proposed algorithm outperforms mainstream image SR algorithms by an average of 0.104,0.224,0.146 dB in peak signal to noise ratio(PSNR),and 0.0349,0.0276,and 0.0181 in structural similarity index measure(SSIM)at magnifications of two,three,and four times.This algorithm can more effectively utilize the original image information,and the reconstructed image has certain improvemenTSin edge and texture details.