Research On Image Compressed Sensing Reconstruction Technology Based On Deep Learning Theory

A important problem in digital media era is data sampling.Traditional Nyquist sampling law requests that the sampling frequency is not less than 2 times the highest frequency of the spectrum of the signal to be sampled.With the era of big data coming,the sampling data and the compression and decompression operations in the later stage are facing a sharp increase in overload.To alleviate these difficulties,the theory of compressed sensing is put forward,it uses special sampling methods to integrate sampling and compression into one process,thus breaking the limitation of Nyquist's sampling law and providing a new set of data sampling theory for various applicatio n scenarios with limited resources.In this paper,the reconstruction theory of image compressed sensing technology is studied.Using the rapid development of deep learning theory as a tool,two deep learning image compressed sensing reconstruction methods are proposed.They are reconstructio n algorithm based on stacked convolutional denoising autoencoder(SCDA)and reconstruction algorithm based on multilayer deconvolution network(CS-De CNN).In SCDA,this paper transforms the denoising autoencoder in the existing SDA reconstructio n algorithm into convolutional denoising autoencoder and convolutional autoencoder which suitable for images,and proposes relevant neural network training methods.In CS-De CNN,the multilevel deconvolution is used instead of the full connection to raise the dimension of the low dimension signal of the compressed sensing.It realizes the step by step and not the one-off dimension,and combines the pixel local adjustment of the ascending dimensio n operation and the decompression process into one process.Thus,the redundant connection of the full connection layer of the neural network is reduced,and the better reconstructio n performance is achieved.Experiments were carried out on Google's open source machine learning platfo rm Tensorflow to deal with image compressed sensing reconstruction problems with SCDA and CS-De CNN methods respectively.The experiment shows that,as an improved version of SDA,SCDA can get better reconstruction precision and reduce space occupation;CS-De CNN can further improve the quality of image reconstruction,and it is the least space occupancy method in several known deep learning image compressed sensing reconstruction methods.In addition,the experiment also shows that,as a deep learning method,the reconfiguration of SCDA and CS-De CNN exceeds the traditional non depth learning methods at low sampling rates,and they have 3 orders of magnitude higher in time performance.