Non-iterative Of Compressed Sensing Algorithm Based On Convolutional Neural Network

Compressed sensing is a reconstruction algorithm that used a small number of measurements to reconstruct signals through nonlinear reconstruction algorithm.At present,the compressed sensing system using iterative optimization algorithm needs complex iterative operations in the process of reconstruction,and the reconstructed image quality is poor in the case of low sampling rate.Convolution neural network can achieve end-to-end non-iterative reconstruction,which can improve the speed of reconstruction.Therefore,we study the image compressed sensing algorithm that based on convolutional neural network,and the specific contents are as follows.Firstly,we apply the convolution neural network to the image compressed sensing.A multi-scale residual reconstruction network is proposed based on the different size of convolution kernel that can extract the characteristics of different scales in the image.And which used a large amount of training data to learn a mapping from measurements to reconstructed images.The experiments show that the reconstruction algorithm based on the multi-scale residual network has better reconstruction results than the iterative optimization algorithm.Secondly,considering that the reconstruction speed of the multi-scale residual reconstruction network is slightly slower than the algorithms of DR~2-Net,we apply the dilate convolution to the reconstruction network,and propose a fast multi-scale residual learning reconstruction network with residual learning.Which reduces the convolution layer of the network,while guarantees the quality of the image reconstruction and improves the speed of the algorithm.At last,the output and measurements of the network are optimized,so that the projection of reconstructed image on the measurement matrix is closer to the measurements.Finally,we propose a reconstruction algorithm based on auto-encode model for compressed sensing,aim to overcome the block effect that caused by the above two algorithms.In order to improve the quality of reconstruction image,we learn the measurement matrix while training the network,and use the acquired measurement matrix to sample the image,reconstruct the image through the auto-encoder model,and then use the pseudo trace network to remove the block effect and artifact in the image.