基于轻量化Transformer的高效图像超分辨率算法研究  

Study of efficient image super-resolution algorithm based on lightweight Transformer

作　　者：高翔 王凡[1] 胡小鹏[1] GAO Xiang;WANG Fan;HU Xiaopeng(School of Computer Science and Technology,Dalian University of Technology,Dalian 116024,China)

机构地区：[1]大连理工大学计算机科学与技术学院,辽宁大连116024

出　　处：《大连理工大学学报》2025年第2期212-220,共9页Journal of Dalian University of Technology

基　　金：国家科技重大专项资助项目(2018YFA0704605)。

摘　　要：基于Transformer的算法在图像超分辨率领域取得的重要性能突破,得益于其捕捉图像中长程依赖关系的强大能力.然而,繁重的计算成本和高GPU显存消耗限制了其在实际中的应用,于是提出了一种基于轻量化Transformer的高效图像超分辨率算法——LISRFormer.该算法引入轻量化Transformer,在捕捉长程依赖关系的同时将复杂度从现有的二次方降为线性.通过跨通道计算交叉协方差,得到可应用于大尺寸图像的转置注意力图.层归一化仅作用于查询和键分支,以保留重要的输入特征.此外,还设计了一种高效门控深度卷积前馈网络(EGDFN),作为Transformer中的前馈网络,进一步恢复准确的纹理信息.在基准数据集上进行的大量定量和定性实验表明,该算法在计算成本和图像重建质量方面优于现有轻量化图像超分辨率算法.Transformer-based algorithms have achieved significant performance breakthroughs in image super-resolution field by the strong ability to capture long-range dependencies in images.However,the heavy computational costs and high GPU video memory consumption limit their practical applications,then an efficient image super-resolution algorithm based on lightweight Transformer(LISRFormer)is proposed.A lightweight Transformer is introduced to this algorithm to capture long-range dependencies while reducing the complexity from existing quadratic to linear.By calculating the cross covariance across channels,a transposed attention map that can be applied to large-sized images is obtained.The layer normalization only affects the query and key branches to preserve essential input features.Moreover,an efficient gated depth-wise-convolution feed-forward network(EGDFN)is designed as the feed-forward network in Transformer to further restore accurate texture information.Numerous quantitative and qualitative experiments conducted on benchmark datasets show that this algorithm outperforms existing lightweight image super-resolution algorithms in terms of computational cost and image reconstruction quality.

关 键 词：图像超分辨率 TRANSFORMER 轻量化 注意力 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]