基于多尺度注意力机制的单幅图像超分辨率重建
Single image super-resolution reconstruction based on multi-scale attention mechanism
阿火黄军 1严华1
作者信息
- 1. 四川大学电子信息学院,成都 610065
- 折叠
摘要
近年来,深度卷积神经网络(CNN)在单幅图像超分辨率重建中取得了明显的进展.在此基础上,提出了一个校准多尺度通道空间注意网络(CMCSAN).CMCSAN由两个关键模块组成:校准多尺度模块(CMSM)和通道空间注意模块(CSAM).CMSM从不同尺度提取特征信息,自适应调整信息特征;CSAM模块可以自动鉴别不同通道的特征信息,有效调整空间的位置权重.实验结果表明,CMCSAN显著增强了挖掘中间特征信息的能力,在单幅图像超分辨率重建中表现出良好的性能.
Abstract
In recent years,deep convolutional neural network(CNN)has made significant progress in single-image super-reso-lution reconstruction.Building upon this foundation,a Calibrated Multi-scale Channel-Space Attention Network(CMCSAN)is in-troduced.CMCSAN comprises two crucial modules:the Calibrated Multi-scale Module(CMSM)and the Channel-Space Attention Module(CSAM).CMSM extracts features from various scales and adaptively adjusts feature information.The CSAM module auto-matically distinguishes different channel features and effectively adjusts spatial positional weights.Experimental results demon-strate that CMCSAN significantly enhances the ability to extract intermediate features and exhibits promising performance in single-image super-resolution reconstruction.
关键词
单幅图像/校准多尺度模块/通道空间注意模块/超分辨率重建Key words
single-image/calibrated multi-scale module/channel-space attention module/super-resolution reconstruction引用本文复制引用
出版年
2024
NETL
NSTL
万方数据