Image Compressed Sensing Based On Neural Network And Reconstruction Algorithm Research

Compressed sensing theory breaks through the limitations of Nyquist sampling theorem,and the original signal can be reconstructed using a small number of projected measurements.At the same time,combining the process of signal sampling and compression has the advantages of saving transmission bandwidth,avoiding resource waste and shortening time duration.Once the theory is put forward,it brings a breakthrough to image reconstruction and has a wide application in remote sensing imaging,image video monitoring,radar imaging and other fields.However,most of the existing compressed sensing image reconstruction algorithms have two shortcomings.On the one hand,in the sampling stage,the matrix is??used to obtain the measurement value,and the original image needs to be divided into blocks,and the final reconstructed image is combined into the final reconstructed image by arranging and splicing the reconstructed small image blocks.As a result,the reconstructed image suffers from severe block effect;on the other hand,the network model in the reconstruction stage has a single structure,and the extracted image features are limited.The network reconstruction accuracy is not high,especially in the areas with rich image details and complex texture features,the reconstruction results are too blurry.In view of the above problems,the research contents of this thesis are as follows:(1)Aiming at the block effect of reconstructed image,a convolutional neural network-based image compressed sensing global measurement network GMNet is proposed.The network model is composed of a sampling sub-network and a fully convolutional residual reconstruction sub-network.The sampling sub-network includes a convolution layer and a deconvolution layer.The convolution layer replaces the traditional measurement matrix and performs global compressed sensing sampling on the complete original image;the deconvolution layer replaces the fully connected layer and deconvolutes the sampled measurement values.,to obtain a low-quality preliminary reconstructed image.The fully convolutional residual reconstruction sub-network consists of three residual blocks,which optimize the output of the final reconstructed image by comparing the original image label and the residual information of each reconstructed image.The experimental results show that the proposed network can avoid the problem of reconstructed image block effect in subjective vision,and can still recognize the image content at a lower sampling rate.(2)On the basis of the image global measurement network GMNet,the reconstruction stage is optimized,and an image reconstruction network MSRA-Net based on the multi-scale residual attention mechanism is proposed.The network includes a sampling sub-network,a multi-scale residual block and a channel attention mechanism module.The multi-scale residual block extracts the multi-scale features of the image by setting convolution kernels of different sizes to improve the reconstruction accuracy;the channel attention mechanism module enables the network to have a global receptive field through the process of compression,excitation and scaling.Adaptively adjust the weights of feature channels.To achieve the purpose of enhancing important feature channels and suppressing redundant feature channels during training.The representative images are selected from the public image dataset for experiments,and the results show that the overall visual effect of the reconstructed images is significantly improved.The peak signal-to-noise ratio in objective indicators is improved by 2-3d B,and the structural similarity has a significant advantage in comparison.To sum up,in this thesis,aiming at the block effect of reconstructed image two major problems in the research field,the two stages of sampling and reconstruction are optimized respectively,and the methods of global measurement and multi-scale residual attention mechanism are proposed.First,the whole image is directly sampled globally,without the need to block the image;then,the multi-scale features of the image are extracted in the reconstruction stage to improve the reconstruction accuracy,and the channel attention mechanism focuses on the areas with rich details and complex texture features.The experimental results show that the reconstructed image by the proposed network has no block effect,and the overall visual effect is improved.In the comparative experiments,both the peak signal-to-noise ratio and the structural similarity of the proposed network have significant advantages.