Compressed Sensing Reconstruction Algorithm Based On Convolutional Neural Network

In this society,due to the rapid development of information,people's demand for image processing and video processing is increasing.Nyquist's sampling law can no longer meet the increasingly higher requirements of image processing in this society due to its large sampling cost,redundant data and waste of a lot of hardware resources.Compressed sensing theory is widely used in image processing because it acquires measured values and restores images at a sub-Nyquist rate.Although the traditional compressed sensing reconstruction algorithm solves the defects of image reconstruction to a certain extent,such as poor anti-interference ability,high complexity,and waste of resources,the traditional compressed sensing reconstruction algorithm also faces some problems: Most of the traditional compressed sensing reconstruction algorithms are iterative,so these reconstruction algorithms have the problems of high computational cost and long reconstruction time.In recent years,with the continuous development of deep learning,convolutional neural networks have entered people's vision.Convolutional neural networks have become another research hotspot due to their advantages such as strong adaptive ability,automatic feature extraction,local perception and weight sharing.Applying the convolutional neural network to the compressed sensing reconstruction algorithm can effectively avoid the shortcomings of the large calculation amount of the traditional reconstruction algorithm,and construct a good reconstruction image to speed up the reconstruction time and achieve the effect of real-time reconstruction.This paper mainly aims at improving the compressed sensing reconstruction algorithm based on convolutional neural network.The main work is as follows:(1)A new compressed sensing reconstruction network CombNet is proposed.It consists of a linear mapping network and a twelve-layer fully convolutional network.The linear mapping network is represented by a fully connected layer.The experiment firstly divides an image into blocks,and divides them into non-overlapping image blocks.Secondly,through compressed sensing technology,the compressed sensing measurement values of these image blocks are used as the input of CombNet,and then through linear mapping network and full convolution The network outputs reconstructed image blocks,and integrates these image blocks to obtain an image with block artifacts,and then passes this image through a filter to obtain the final reconstructed image.The linear mapping network in CombNet restores the image information.The fully convolutional network enhances CombNet learning ability and further improves the reconstruction accuracy of CombNet.The experiment compares the reconstruction peak signal-to-noise ratio of CombNet with other algorithms.The experiment proves that CombNet has a higher weight than the traditional reconstruction algorithm and three existing convolutional neural network-based compressed sensing reconstructionalgorithms.Structural accuracy and structural similarity.(2)Based on CombNet,the experiment adds a residual block to enhance the learning ability of the CombNet network,which is called FCNet.In the experiment,the standard test images in the Berkeley image library were selected and reconstructed at four different sampling rates.By comparing the peak signal-to-noise ratio of the seven algorithms with and without filters,the experimental conclusions were drawn.Experiments prove that FCNet has a smaller reconstruction error at the same sampling rate,can extract scene information more effectively,and improve real-time reconstruction while improving image reconstruction accuracy and achieving better reconstruction visual effects.