Research On Compressed Sensing Algorithm Based On Deep Learning

With the rapid development of communication technology and the Internet,the era of big data has arrived,and the efficiency of information transmission has become more and more prominent.Compressed Sensing(CS)theory has brought tremendous changes to the field of information transmission.It also realizes the signal sampling and compression process,and can use fewer observations to achieve accurate signal recovery,which is better and faster It provides a solution for transferring massive amounts of data.The current compressed sensing theory mainly faces two major problems:how to improve the reconstruction quality at the decoding end,and how to design the sampling matrix at the encoding end.Therefore,based on the deep learning method,this paper studies the above two issues.The main work and achievements include:(1)Propose a block-based compressed sensing algorithm based on deep learning.At the decoding end,the reconstruction effect of the traditional compressed sensing reconstruction algorithm based on iterative optimization is not ideal,and the algorithm has high computational complexity and long reconstruction time.To solve the above problems,the model adopts a three-segment network structure of sampling module-initial reconstruction module-depth reconstruction module.A full-precision convolutional layer without bias and activation is used as the sampling module.The initial reconstruction module uses the depth separable convolution to obtain the information in each block while also learning the information between each block to reduce the blocking effect.The deep reconstruction module combines channel splicing and dilation convolution,and designs an intermediate layer result for channel splicing and then input to the residual block of the lower layer.The residual block is stacked and dilated convolution is added between the residual blocks to improve the reconstruction quality.Experiments show that compared with other deep compressed sensing models,this model has certain advantages in objective indicators and subjective experience,and shows that the details of the reconstructed image change as the sampling rate increases.(2)Propose a ternary method of sampling matrix based on attention mechanism.At the encoding end,the common sampling matrices are mostly random matrices,which have nothing to do with the signal being sampled,and cannot fully express the structural characteristics of the signal,and are stored as floating-point numbers,which often occupies more storage space.In response to the problems,a compressed sensing framework called ATP-Net(Attention-based ternary projection network)is realized through the ternary sampling matrix and the decoding end.The main idea of the ternary method(-1,0,+1)is to use the attention mechanism to evaluate the importance of the weights after the sampling matrix is binarized(-1,+1),and then compare the importance the low weight is removed to achieve ternary.The sampling matrix of ATP-Net is based on the convolutional layer,which can be obtained adaptively through data drive without manual intervention,and the ternary matrix elements only need 2bit(sign bit and value bit)storage,which can improve the hardware application of the sampling matrix.Through experiments and comparison with common compressed sensing algorithms,it is shown that the image reconstruction quality of ATP-Net can still maintain a good reconstruction effect on the basis of the sampling matrix tri-value.(3)Designed and implemented a set of compressed sensing framework based on deep learning.Written and implemented a compressed sensing simulation system.The system includes modules such as preprocessing,compression sampling at the encoding end and reconstruction at the decoding end.Starting from the size of the memory occupied by the input data and the sampled data,the simulation results show that the compressed sensing algorithm proposed in this paper can indeed achieve compressed sampling at the corresponding sampling rate and has a good reconstruction quality.