Research On Image And Video Reconstruction Algorithm Based On Compressive Sensing

Compressive sensing theory has broken the restriction of Nyquist sampling theorem on sampling frequency,which made it widely used in various fields.In image processing,combining the compressive sensing and image can not only reduce the cost of data compression and sampling,but also improve the efficiency of image processing.Therefore,image reconstruction is one of the cores of compressive sensing research,whose purpose is to restore the original image from little measurements.However,the image reconstruction itself is a NP-Hard problem,how to use the prior to reconstruct the image efficiently is still a focus of current researches.With different priors,researchers have proposed amounts of image reconstruction algorithms,but the performance of these algorithms are not same and satisfactory.On this basis,this paper takes image and video as research objects and proposes the novel image and video reconstruction algorithms with better performance.Firstly,this paper studied the traditional image reconstruction algorithms.Starting with image prior,this paper proposed the reweighted double sparse constraint image reconstruction algorithm.That is,the algorithm combines the signal sparsity and residual constraint based on non-local similarity,then establish the model by the means of reweighting,finally splits the model into sub-problems under the framework of the split Bregman iteration algorithm and solve them in the ways of iteration.Experiments performed by the algorithm is verified,at the ratio of 0.2,the average PSNR of the images reconstructed by proposed algorithm compared to the Co S algorithm and the MH algorithm has increased by 2.0 d B and 1.5 d B respectively,which effectively improves the quality of the reconstructed image.Current reconstruction algorithms mostly use two-dimensional images as study objects,but less focus on video.Therefore,this paper proposed the image and video reconstruction algorithm based on tensor approximation and spatio-temporal correlation,which uses higher singular value decomposition to realize the low-rank approximation of tensor.That is,based on the non-local similarity,the similar patches in multi-frames are formed to tensor,and then the low-rank approximation is used to reconstruct the tensor effectively.The proposed reconstruction model is solved in the ways of iteration under the framework of alternating direction method of multipliers.On this basis,a video reconstruction model is further proposed by combining the spatio-temporal correlation.Finally,the performance of the algorithm is verified by simulation experiments.Under the three video sequences,the average PSNR of the video reconstructed by the algorithm is increased by 3.31 d B compared with the video MH algorithm.The emergence of deep learning provides a new way for the development of compressive sensing theory.Meanwhile,the reconstruction algorithm based on deep learning also shows great advantages in performance of reconstruction model and quality of restored image.In the framework of deep learning,this paper proposed a deep compressive sensing reconstruction model based on non-local priors.The model is first established by combining image sparsity and non-local similarity,and then it is decomposed into multiple sub-problems by using half-quadratic splitting.Finally,each sub-problem is solved by deep learning,and a complete end-to-end trainable image reconstruction network is established jointly.The performance of the proposed algorithm is verified by experimental simulation.At the rate of 0.2,the PSNR value of the reconstructed image of the proposed algorithm has increased 1.98 d B,1.81 d B and 0.24 d B compared with the image reconstruction networks ISTA-Net,CSNet and SCSNet based on deep learning.