Research On Image Compressed Sensing Algorithm Based On Dilated Convolution

With the development of society,people ’s requirements for data are also increasing,resulting in a corresponding increase in the speed of information transmission.In the process of data collection and processing,the demand for acquisition frequency and processing speed is also increasing.Due to redundancy and sparseness,images can be reconstructed according to compressed sensing theory.The goal of image compressed sensing is to accurately reconstruct the image from less than the measured value required by the Nyquist sampling theorem.To achieve image compressed sensing,two key problems need to be solved.One is how to design an efficient sampling matrix to obtain the measured value,and the other is how to quickly recover the highquality reconstructed image from the measured value.At present,the use of convolutional neural networks to complete the task of image compressed sensing reconstruction has achieved good results.Such networks are usually composed of sampling networks,preliminary reconstruction networks,and deep reconstruction networks.However,the existing image compressed sensing reconstruction models generally have the problems of high computational complexity and sacrificing reconstruction time in exchange for reconstructed image quality in the design of reconstruction network.Therefore,the focus of this paper is mainly on the reconstruction network,aiming to propose a deep reconstruction network that takes into account the reconstruction quality and running speed to solve the above problems.The specific contents are as follows:(1)A feature aggregation residual block is proposed,and an image compressed sensing model based on feature aggregation residual block is constructed,which is called TDCN＿FARB.The residual block cascades a combination layer composed of a combination layer composed of dilated convolution with dilated rate of 1,batch normalization layer and Re LU activation function,a combination layer composed of dilated convolution with expansion rate of 2,batch normalization layer and Re LU activation function,and a combination layer composed of Kronecker convolution with expansion rate of 3,batch normalization layer and Re LU activation function.Then,the output of each combination layer is fused by feature fusion operation.Finally,the input and output are connected to form a residual form,and the feature aggregation residual block is obtained.The experimental results show that TDCN＿FARB is superior to most of the mainstream algorithms in recent years.(2)A tree-structured residual block is proposed,and an image compressed sensing model based on tree-structured residual block is constructed,which is called TDCN＿TSRB.The residual block cascades three dilated convolutions with expansion rates of 1,2 and 3 respectively,and then the output of each dilated convolution is also fused by feature fusion operation.Finally,the input and output are connected to form a residual form to obtain a tree-structured residual block.Compared with the feature aggregation residual block,the batch normalization layer and pooling layer used in it are removed.The experimental results show that the proposed TDCN＿TSRB model is superior to TDCN＿FARB and most algorithms in recent years in terms of image compressive sensing reconstruction performance.