Image Compression Sensing Reconstruction Based On Convolutional Neural Networks

Due to the increasing demand for information,people have increasingly high requirements for sampling methods,processing speed,storage space and information quality in the process of data acquisition.The introduction of Compressive Sensing(CS)theory makes it possible for the measurement rate to be less than twice the signal bandwidth.And this theory can achieve signal compression while sampling signals,which greatly saves storage space.At the same time,with the development of neural networks,a large number of image compression perceptual reconstruction algorithms based on Convolutional Neural Networks(CNN)have been proposed.However,these algorithms still have two major problems.One is that the method of block measurement and block by block reconstruction of the input image can destroy the structural information of the image,which causes serious block effects in the reconstructed image.Secondly,in the reconstruction section,in order to improve the clarity of the reconstructed image,the algorithm generally adopts the method of stacking multiple convolutional layers,which leads to higher computational complexity and longer reconstruction time.Therefore,this paper conducts research on the above two major issues,and the main contents are as follows:(1)An image compressed sensing reconstruction network based on inverse residual blocks(IRB-CS)is proposed.IRB-CS network will solve the following problems: the blocking effect introduced by image segmentation and measurement,high computational complexity and long reconstruction time.IRB-CS is composed of a three-segment network structure of sampling network,initial reconstruction network and depth reconstruction network.The sampling network does not perform a blocking operation on the input image,but instead uses full convolution to measure the entire image,thereby solving the blocking effect problem.In the initial reconstruction network,deconvolution is used to restore the initial reconstruction image.The depth reconstruction network is composed of the inverse residual block path,the residual block path and the Global Attention Mechanism(GAM).The inverse residual block performs feature extraction from both depth convolution and point-by-point convolution,which validly decreases the amount of arguments and decreases the computational complexity.Due to the lack of attention to global information in the process of sampling and reconstruction,the global attention mechanism is introduced,which includes spatial attention and channel attention.GAM not only reduces the information loss,but also abstracts the information of the entire image.Test results show that compared with other networks,IRB-CS can reconstruct clearer images at low measurement rate,cut down block effects and shorten reconstruction time.(2)This paper proposes an image compression perceptive reconstruction network based on self-attention mechanism(SAMNet),which starts with the design of compression sampling structure and the improvement of reconstruction network.In view of the block effect caused by measuring and reconstructing block by block,and then splicing the reconstructed image blocks,the network model designs a convolution self-attention structure.The output of the self-attention mechanism integrates each input feature,so that the compressed sensing measurement can extract more rich features.At the same time,in the reconstruction part,all the measured values of the whole image are used to reconstruct the image,which effectively cuts down the block effect.In order to address the issues of high computational complexity and long reconstruction time in the reconstruction network,SAMNet introduced dense blocks and Bottleneck Transformers(Bo TNet)into the reconstruction network.The feature reuse feature of dense blocks allows SAMNet to reduce the amount of network parameters and shorten the reconstruction time.The bottleneck transformer extracts all the information of the image and can obtain the reconstructed image with higher definition.The test results display that under the Set5 dataset,when the Measurement Rates(MR)are 0.01,0.04,0.10 and 0.25 respectively,the average Peak Signal to Noise Ratio(PSNR)of the algorithm is 0.2-1.3 d B higher than that of CSNet+,therefore,SAMNet can reconstruct clearer images.