基于残差学习的多阶段图像压缩感知神经网络

Multistage Image Compressive Sensing Neural Network Based on Residual Learning

在传统的图像压缩感知研究中存在两个主要问题:在采样端,传统的线性采样方式存在一定程度的局限性;在重构端,传统的优化重构算法有着高耗时的缺陷。新发展的图像压缩感知神经网络框架虽然很好地解决了重构速度问题,但重构效果一般。为了解决上述问题,文中提出一种基于残差学习的多阶段图像压缩感知网络——MSResICS,其中具体包含3个子网络:采样子网络,初始重构子网络与图像增强子网络。在观测端,提出的非线性采样子网络,利用残差学习打破了传统采样方法的局限性,在采样值中保留了更加丰富的图像信息;在重构端,通过引入插值卷积,初始重构子网络从采样值中进行特征整合以获取质量优良的初始重构图像,最终利用残差学习与插值卷积,提出多阶段图像增强子网络以进一步细化重建图像,提升最终效果。大量仿真实验结果表明,与现有最优的图像压缩感知重构算法相比,MSResICS拥有更加优良的图像压缩感知重构精度。

There are two main problems in traditional Image Compressive Sensing(ICS):in sampling aspect,traditional linear sampling methods have some limitations;in reconstruction aspect,optimization-based reconstruction methods are highly time-consuming.Newly proposed ICS Neural Network can successfully deal with the speed problem in reconstruction,but lacks the accuracy of traditional algorithms.To solve this problem,a novel multistage ICS network based on residual learning(MSResICS)was proposed.It consists of three sub-networks,namely,sampling sub-network,initial reconstruction sub-network,and image enhancement sub-network.In sampling stage,with the help of residual learning,a nonlinear sampling sub-network was designed,which breaks the limitation of conventional sampling method and retains richer image information in samples.In reconstruction stage,the initial reconstruction sub-network extracts features from samples and obtain an initial reconstructed image of high quality by introducing interpolation convolution.With residual learning and interpolation convolution,a multistage image enhancing sub-network was proposed to further refine the reconstruction image and improve the quality of final result.Extensive simulations show that MSResICS has a better reconstruction accuracy than the existing optimal ICS reconstruction methods.