电子学报2023，Vol.51Issue(10)：2812-2820,9.DOI:10.12263/DZXB.20221420
频率引导的双稀疏自注意力单图像去雨算法
Frequency-guided Dual Sparse Self-Attention Algorithm for Single Image Deraining
摘要
Abstract
Existing Transformer-based algorithms for single image deraining achieve state-of-the-art performance but leading to reasonable computational loads while failing to effectively process real-world rainy images.To this end,we pro-pose a frequency-guided dual sparse self-attention Transformer for single image deraining(FDSATFormer).Initially,our proposed method utilizes the spatial sparse factor and the channel reduction factor to extract accurate global information and significantly decreases the amount of computation.Furthermore,we present dual sparse self-attention feature learning net-work(DSFL)to deal with the problem that Transformer is difficult to represent self-attention on high-resolution feature maps.Meanwhile,by exploring the spectral changes of rainy image before and after removing rain streaks,we develop a frequency-guided feature enhancement module(FFE),which exploits the global information from the frequency domain to guide the accurate learning of spatial features in network encoders.In addition,the encoder and decoder of most existing methods follow the similar principles,resulting in almost double computational burden.To handle with this issue,we pro-pose a hierarchical feature decoding and reconstructing network(HFDR),which uses non-parametric spatial neighborhood shift(SNS)to construct the feature decoding network and achieves fine results while further reducing the overall computing burden.Experimental results show that,our method improves the average peak signal noise ratio by 3.13 dB and 0.12 dB,and achieves performance gains of 1.39％and 1.07％in average structure similarity over the state-of-the-art Uformer and Restormer.关键词
计算机视觉/图像去雨/自注意力网络/稀疏注意力/空间移位/频率引导学习Key words
computer vision/image deraining/Transformer/sparse attention/spatial shift/frequency-guided learning分类
信息技术与安全科学引用本文复制引用
文渊博,高涛,陈婷,张千禧..频率引导的双稀疏自注意力单图像去雨算法[J].电子学报,2023,51(10):2812-2820,9.基金项目
国家自然科学基金(No.52172379,No.62001058) （No.52172379,No.62001058）
长安大学中央高校基本科研业务费专项资金资助项目(No.300102242901)National Natural Science Foundation of China(No.52172379,No.62001058) （No.300102242901）
Fundamental Re-search Funds for the Central Universities(No.300102242901) （No.300102242901）
