Compressed Sensing Image Reconstruction Based on Convolutional Neural Network

Abstract

Compressed sensing theory is widely used in image and video signal processing because of its low coding complexity, resource saving, and strong anti-interference ability. Although the compression sensing theory solves the problems brought by the traditional signal processing methods to a certain extent, it also encounters some new problems: the reconstruction time is long and the algorithm complexity is high. In order to solve these problems and further improve the quality of image processing, a new convolutional neural network structure CombNet is proposed, which uses the measured value of compression sensing as the input of the convolutional neural network, and connects a complete connection layer to get the final Output. Experiments show that CombNet has lower complexity and better recovery performance. At the same sampling rate, the peak signal-to-noise ratio (PSNR) is 12.79-52.67 higher than Tval3 PSNR, 16.31%-158.37% higher than DAMP , 1.00%-3.79% higher than DR²-Net, and 0.06%-2.60% higher than FCMN. It still has good visual appeal when the sampling rate is very low (0.01).

High-speed computational ghost imaging with compressed sensing based on a convolutional neural network

“…The quantitative results (Fig. 5) show that the PSNR of CGI with CS-CNN is average 7% higher than that of CGI with DL under the same reconstructed frame number, SSIM increased by 12% on average [26]. Consequently, under the condition of producing the same image quality, CS-CNN has the faster imaging speed.…”

Deep Learning Techniques for Compressive Sensing-Based Reconstruction and Inference -- A Ubiquitous Systems Perspective

“…Thus, as the signal size grows, so does the network, imposing a large computational complexity on the training algorithm and risking potential overfitting. The solution proposed in [104] and adopted by similar approaches [69,81] is to divide the signal into smaller blocks and then sense/reconstruct each block separately. From the reconstruction time point of view, the simulation results show that this approach beats the other methods, whereas the of quality of the reconstruction does not necessarily overshadow that of other state-of-the-art recovery algorithms.…”

“…One of its drawbacks is the fact that this approach uses a blocky measurement matrix, to reduce the network complexity and hence, the training time, and therefore, ReconNet does not exploit potentially strong dependencies that may exist between the reconstructions of different blocks. CombNet provides improved quality of reconstruction by using a deeper network structure and a smaller convolution core [81].…”

“…Going a step further from these block based approaches ( [69,81]), the work in [102] introduces the first DL framework that works for images acquired with measurement matrices of arbitrary sizes, and not only with blocky matrices. This network, named DeepInverse, is a CNN architecture, with a fully connected linear layer, whose weights are designed to implement the adjoint operator (transpose) of the measurement matrix.…”

“…• minimize number of parameters; • reduce memory storage; • increase learning speed; • can better handle larger inputs; • capture structure in images; DeepInverse [102], images, Gaussian [81], images, Gaussian KCSNet [18], images, learned WDLReconNet [88],images, random [139], MRI, 2D Poisson [103], generic, learned [119], MRI, Cartesian [89], image, learned MLP/FC [2], images, learned…”