基于混合注意力机制的图像超分辨重建算法  

Image Super-Resolution Reconstruction Based on Hybrid Attention Mechanism

作　　者：李孟歆[1] 贾欣润 李松昂 LI Meng-xin;JIA Xin-run;LI Song-ang(School of Electrical and Control Engineering,Shenyang Jianzhu University,Shenyang Liaoning110000,China)

机构地区：[1]沈阳建筑大学电气与控制工程学院,辽宁沈阳110000

出　　处：《计算机仿真》2023年第12期236-241,共6页Computer Simulation

基　　金：国家自然科学基金项目资助(62133014)。

摘　　要：现有的大多数基于深度学习的单幅图像超分辨率算法,是通过深化和拓宽网络结构来提取特征,而对于信息占比不同的空间域和通道域没有作区分,导致大量计算资源的浪费。针对上述问题,提出了一种通道-空间混合注意力模块,通过捕捉通道域和空间域内重要性的差异从而更高效地分配计算资源,以加快网络收敛,提高网络性能。采用跳跃连接的方式融合全局特征,加强网络内信息的传递,使得分层信息被充分利用。同时在网络中引入密集连接网络,以做到特征的复用,加强信息的传输。实验结果表明,上述算法在客观指标评价和主观视觉效果方面均优于比较算法。Most of existing single image super-resolution algorithms based on deep learning employ features by deepening and widening the network structure.However,the lack of distinction between spatial and channel domains with different information ratio results in an excessive waste of computing resources.To solve this problem,this paper proposes a channel-spatial hybrid attention module,it allocates computing resources more efficiently by capturing the difference in the importance of each channel and spatial domain information to speed up network convergence and to improve network performance.The global features are fused by skip connections to enhance the flow of information in the network,so the hierarchical information can be fully utilized.By simultaneously introducing a densely connected convolution network into the network,it can achieve characteristics of multiplexing and strengthening information transmission.The experimental results show that this proposed algorithm exceeds comparison algorithms in both objec⁃tive index evaluation and subjective visual effects.

关 键 词：单幅图像超分辨 密集连接网络 残差连接 通道注意力机制 空间注意力机制 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]