基于SPL迭代思想的图像压缩感知重构神经网络

Image Compressive Sensing Reconstruction Network Based on Iterative SPL Theory

由于神经网络强大的学习能力与快速的运行速度,近年来基于深度学习的图像压缩感知(Image Compressive Sensing,ICS)研究备受关注.然而,大多数现有ICS神经网络的结构设计忽略了传统迭代重构算法中的数学理论基础,无法有效利用信号中的先验结构知识,可解释性较差.为了保留优化算法核心思想并同时利用深度学习的高性能,本文使用可学习的卷积层替代了传统平滑投影Landweber算法(Smooth Projected Landweber,SPL)中的人工设计参数,提出一种新型ICS神经网络SPLNet.在SPLNet中,设计了一个独特的网络结构SPLBlock实现SPL迭代过程中的三个核心步骤:(1)去除块效应的维纳滤波器;(2)在凸投影集合上的近似操作;(3)实现稀疏表示及去噪的变换域双变量收缩.仿真实验结果表明:与现有最优的ICS优化迭代算法GSR相比,SPLNet的重构图像平均PSNR提升了0.78dB;与最优的神经网络框架SCSNet相比,SPLNet的重构图像平均PSNR提升了0.92dB.

Due to its great learning ability and fast processing speed,deep learning-based image compressive sensing(ICS)methods attract a lot of attention in recent years.However,the design of most existing ICS neural networks architecture ignore the mathematical theory in iterative optimization-based methods and cannot effectively use the prior structure knowledge in the signal,leading to lack of the interpretability.In order to retain the core ideas of the optimization algorithm and utilize the high performance of deep learning,this paper uses learnable convolutional layers to replace the predefined filters and artificial design parameters in the traditional smooth projected Landweber algorithm(SPL),and proposes a ICS neural network named SPLNet.In SPLNet,we design a unique network structure SPLBlock to implement three key steps in SPL iteration:(1)Wiener filter for removal of blocking artifacts;(2)approximation with projection onto the convex set;(3)bivariate shrinkage on transform domain for sparse representation and denoising.Experimental results indicate that,compared with current state-of-the-art ICS optimization iterative algorithm GSR,the average reconstructed image PSNR of SPLNet are improved by 0.78dB,and compared with state-of-the-art neural network framework SCSNet,the average reconstructed image PSNR of SPLNet are improved by 0.92dB.