基于多尺度残差网络的全局图像压缩感知重构

Global Image Compressed Sensing Reconstruction Based on Multi-scale Residual Network

压缩感知技术突破了奈奎斯特采样定理的限制,但目前压缩感知的传统迭代重构思路已经陷入了瓶颈,例如TV算法、DAMP算法、NLR-CS算法,存在着重构质量差或重构速度慢的问题。最近,深度学习的方法在传统的压缩感知图像重建中取得了显著的进步。提出了一个基于多尺度残差网络的图像CS重构模型,该模型由多个多尺度残差块构成。对于每个多尺度残差块,其具有多条旁路,每个旁路应用不同大小的卷积核自适应检测不同尺度的图像特征,并且每个多尺度残差块的输出都用于全局的特征的分层融合,从而使图像恢复更高的质量。从实验室结果可以看出,所提出的方法在图像的PSNR和SSIM上均优于现有基于深度学习的方法,并且重建速度和质量优于传统的重构方法。

This paper proposes an image CS reconstruction model based on multi-scale residuals network,which is composed of multi-scale residuals.For each multi-scale residual block,it has multiple bypasses,each of which applies convolution kernel of different sizes to adaptively detect image features of different scales,and the output of each multi-scale residual block is used for hierarchical fusion of global features,so as to restore higher image quality.It can be seen from the laboratory results that the method proposed in this paper is better than the existing deep learn-based methods in both PSNR and SSIM,and the reconstruction speed and quality are better than the traditional reconstruction methods.