基于整体注意力的图像超分辨率反投影网络-《集美大学学报（自然科学版）》

文章信息/Info

Title:: Image Super-Resolution Back-Projection Network Based on Holistic Attention

Author(s):: QI Tangyao; ZHANG Dongxiao; School of Science,Jimei University,Xiamen 361021,China

Keywords:: super-resolution; back-projection; holistic attention; layer attention; channel spatial attention

摘要:: 针对图像超分辨率重建问题，提出了一种基于整体注意力的反投影网络。该网络中的上下迭代投影模块与整体注意力模块相互协同，共同捕获更多特征信息。〖JP〗通过上下迭代投影模块的相互转化，获取不同分辨率特征之间的关联性，帮助学习到高频纹理细节。使用层注意力模块，将上下投影输出的特征层分别进行层之间的注意力加权，用以寻找不同层、通道和位置之间相互依存的关系。此外，针对最后的上投影模块使用了通道空间注意力模块，可以引导网络学习到通道内部和通道间的相关信息。消融实验表明，两种注意力模块对特征信息的增强有助于提升重建效果。对比实验表明，本文提出的基于整体注意力的反投影网络模型具有较好的重建效果。

Abstract:: Aiming at the problem of image super-resolution reconstruction,this paper proposes a back-projection network based on the overall attention.The up-and-down iterative projection module and the overall attention module in the network work together to capture more feature information.Through the mutual conversion of the up and down iterative projection modules,the correlation between different resolution features is obtained,which helps to learn high-frequency texture details.In this paper,the layer attention module is used to weight the attention between the layers of the feature layers output by the upper and lower projections to find the interdependent relationship between different layers,channels and positions.In addition,the spatial channel attention module is used for the final up-projection module,which can guide the network to learn relevant information within and between channels.Ablation experiments show that the enhancement of feature information by the two attention modules helps to improve the reconstruction effect.Comparative experiments show that the back-projection network model based on holistic attention proposed in this paper has a good reconstruction effect.