双U型门控网络融合非局部先验的图像压缩感知重建方法  

Dual U-shaped network with non-local prior fusion for image compressed sensing

作　　者：林乐平[1,2] 胡尚鋆 欧阳宁[1,2] Lin Leping;Hu Shangjun;Ouyang Ning(Key Laboratory of Cognitive Radio&Information Processing Ministry of Education,Guilin University of Electronic Technology,Guilin Guangxi 541004,China;School of Information&Communication,Guilin University of Electronic Technology,Guilin Guangxi 541004,China)

机构地区：[1]桂林电子科技大学认知无线电与信息处理省部共建教育部重点实验室,广西桂林541004 [2]桂林电子科技大学信息与通信学院,广西桂林541004

出　　处：《计算机应用研究》2024年第11期3509-3514,共6页Application Research of Computers

基　　金：国家自然科学基金资助项目(62001133,61661017,61362021);广西科技基地和人才专项资助项目(桂科AD19110060);广西高等学校千名中青年骨干教师培育计划资助项目;广西自然科学基金资助项目(2017GXNSFBA198212);广西无线宽带通信与信号处理重点实验室基金资助项目(GXKL06200114)。

摘　　要：针对目前基于非迭代式网络的图像压缩感知重建方法存在着细节处理能力不足以及测量值利用不充分的问题,提出了一种双U型门控网络(dual U-shaped gated network, DUGN)用于图像压缩感知重建。该方法在原有的U型结构网络的基础上进行了改进,提升了U型结构网络在压缩感知任务中的学习能力。在测量值的利用上,结合交叉注意力机制,提出了一种测量值非局部融合模块(measurements non-local fusion, MNLF),用于将测量值中的非局部信息融合到深层网络中,指导网络进行重建,提升模型性能。此外,在基本模块的设计上,提出了窗口门控网络模块(window gated network, WGN),增强了网络的细节处理能力。实验结果表明,与已有的压缩感知重建方法相比,DUGN在Set11数据集上有着更高的PSNR和SSIM,且在图像重建的真实性上有着更好的表现。Aiming at the problem that insufficient detail processing ability and inadequate use of measurement values in current non-iterative network-based image compressed sensing methods,this paper proposed a novel dual U-shaped gated network(DUGN)for image compressed sensing.This network improved upon the existing U-shaped network structure and enhanced the learning ability in compressed sensing tasks.This paper proposed a measurements non-local fusion(MNLF)module based on cross-attention mechanism,MNLF integrated non-local prior information from measurements into the deep network and guided the network reconstruction.Furthermore,this paper proposed a window gated network(WGN)module as the basic blocks to enable the network to learn larger ranges of image information and enhance the ability to process details.Experimental results demonstrate that compared with existing compressed sensing reconstruction methods,DUGN has higher PSNR and SSIM on Set11 and better performance in the authenticity of image reconstruction.

关 键 词：图像压缩感知重建 非局部先验 U型网络 门控网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]