基于自适应非线性网络的压缩感知重构算法

Compressed Sensing reconstruction algorithm based on depth learning adaptive nonlinear networks

摘要:

针对传统压缩感知(CS)进行复杂的迭代运算，重构时间长且质量差等问题，结合深度学习方法，提出一种自适应非线性测量卷积神经网络(NMECNN)的压缩感知重构算法。本算法将图像整体宽高进行压缩，作为测量网络替代传统的随机测量矩阵进行图像重建，同时利用多个扩张卷积层和上采样PixelShuffle方法获取图像不同尺度细节信息。通过与其他文献进行实验对比，本算法在不同采样率下，平均峰值信噪比(PSNR)分别高于MSRNets算法1 dB,0.7 dB,0.82 dB,1.61 dB；结构相似性(SSIM)值分别高0.03,0.04,0.24,0.10个单位，重构时间在CPU上比MSRNet算法快0.175 5 s, 0.399 8 s,0.41 s,0.396 s。最后通过大数据集与噪声实验，验证了本算法图像重构质量明显提高，重构时间大幅缩短，具有很强的抵抗噪声攻击能力。

Abstract:

Aiming at the complicated iterative operations of traditional Compressed Sensing(CS), long reconstruction time and poor quality, a compressed sensing reconstruction algorithm for Non-linear Measurement Convolutional Neural Network(NMECNN) is proposed by combining the deep learning method. This algorithm compresses the overall width and height of the image as a measurement network to replace the traditional random measurement matrix for image reconstruction. At the same time, it uses multiple expanded convolutional layers and upsampling PixelShuffle methods to obtain detailed information of different scales of the image. Through experimental comparison with other documents, the average Peak Signal to Noise Ratio(PSNR) values of this algorithm at different sampling rates are higher than that of MSRNets algorithm by 1 dB,0.7 dB,0.82 dB,1.61 dB, and the Structural Similarity(SSIM) values are higher by 0.03,0.04,0.24,0.10 units. The reconstruction time in the CPU is less than that of the MSRNet algorithm by 0.175 5 s,0.399 8 s,0.41 s,0.396 s, respectively. Through big data sets and noise experiments, it is verified that the image reconstruction quality of this algorithm is significantly improved, the reconstruction time is greatly shortened, and it has a strong ability to resist noise attacks.