Research On Image Compressed Sensing Based On Deep Learning

As the era of digital information and internet has come,the transmission and storage demand of digital information such as images has drastically increased.In the process of image transmission and storage,the procedure of compressing and sampling is essen-tial.Nyquist Sampling Theory issued that only if the sampling rate is larger than or equal to 2 times of the highest frequency in the frequency spectrum of the original signal will it be reconstructed without distortion,which makes compressing and sampling of high-frequency signal such as images very difficult to realize.However,Compressed Sensing theory has had this difficulty solved.It combines the procedure of compressing and sam-pling together.And it also could compress and sample then precisely reconstruct images under a low sampling rate,which breaks through the limitation of Nyquist Sampling The-ory.In the research of Compressed Sensing,the current research hotspots are designing a adaptive sampling matrix and a reconstruction algorithm that has high performance of reconstruction effect.In the recent year,the combination of rising Deep Learning theory and traditional Compressed Sensing thoery has developed new Compressed Sensing theory which is based on Convolution Nerual Networks?CNN?and has high performance of image recon-struction effect.Under this premise,a Compressed Sensing theory which utilizes CNN adaptive sampling matrix to compress and sample images then utilizes Orthogonal Match-ing Pursuit?OMP?algorithm to reconstruct images and a Compressed Sensing reconstruc-tion algorithm which is based on CNN Residual Networks to reconstruct images has been designed.The main research procedure of this thesis is showing below:1?The producing process of Gussian random sampling matrix,Bernoulli random sampling,Toeplitz sampling matrix and the main procedure of OMP algorithm and the features of MNIST image dataset are expounded.Then the reconstruction effect of im-age compressed sensing frameworks of the combination of these three traditional sam-pling matrices and OMP algorithm is compared.The simulation result shows that the best performance of reconstruction effect is obtained when utilizing the compressed sensing framework of the combination of Gussian random sampling matrix and OMP algorithm.Also,it shows that a logical designing defect of OMP algorithm is discovered in the sim-ulation,which makes the performance of reconstruction effect of some of the MNIST images poor.2?A compressed sensing framework which combines an adaptive sampling ma-trix which can change as the features of original images change by utilizing CNN with OMP algorithm?CNN-OMP?is designed.Then the reconstruction effect of CNN-OMP and compressed sensing frameworks of the combination of these three traditional sam-pling matrices and OMP algorithm is compared by utilizing TensorFlow framework in simulation.The simulation result shows that the reconstruction error rate of CNN-OMP compared to that of these three traditional sampling matrices dramatically decreases by about 60%under a low sampling rate.Also,it shows that this adaptive sampling matrix can improve the poor performance of reconstruction effect of some of the MNIST images which is caused by the logical designing defect of OMP algorithm.3?An end-to-end image compressed sensing framework which combines a recon-struction algorithm which is based on CNN and Residual Networks with CNN adaptive sampling matrix?CNN-ResNet?is designed.Then the reconstruction effect of it and CNN-OMP is compared by utilizing Pytorch framework in simulation.The simulation result shows that the reconstruction error rate of CNN-ResNet is¹₂₀of that of CNN-OMP.Also,it shows that the logical designing defect of OMP algorithm can be avoided when utiliz-ing CNN-ResNet due to its end-to-end feature,which makes it can precisely reconstruct almost all images in MNIST image dataset.