Research On Video Recovery Method Based On Tensor Analysis

With the development of compressive sensing,more and more video compressive systems have been proposed.Among them,the representative is the snapshot compressive imaging system,which captures high-dimensional video or image data by low-dimensional sensors and outputs a single measurement that is much less than the amount of raw original data.Therefore,it reduces the required transmission bandwidth and hardware storage.The question is how to accurately recover the original video based on the measurements obtained? This can be solved by the decoding module of the snapshot compressive imaging system.Unfortunately,most existing decoding methods achieve poor reconstruction quality,strangling the wide application of this compressive technology.After research,we found that most video compressive imaging system decoding algorithms use matrix and vector to represent video data,which often loses the space-time structure information of the original video.Considering that a gray video is composed of three dimensions,i.e.,width,height,and time,thus the video with multiple frames can be represented as a third-order tensor.Then,exploiting the corresponding tensor based algorithm to decode compressive measurement,it can preserve the space-time information of the original video as much as possible,thereby improving the quality of video recovery.In this thesis,t-product based tensor model is applied to study the video decoding problem of snapshot compressive imaging system.The main tasks are:First,the t-product based tensor model is studied in depth.Starting from the basic concepts and definitions,the basic characteristics of the tensor model based on t-product are studied,including tensor transpose,tensor orthogonality,and the computing process of tensor singular value decomposition.Moreover,the snapshot compressive imaging system is described in detail and the existing video decoding algorithms are studied in depth.Secondly,the video decoding problem is modeled and analyzed by tensor algebra operation.Based on the nonlocal similarity of video data,an alternative optimal decoding algorithm based on nonlocal tensor recovery is proposed.This algorithm decomposes video decoding problem into two subproblems: similar block grouping and tensor recovery.In order to solve the tensor recovery problem,we first propose a tensor optimal approximation theorem based on t-product.Then,exploiting the signal estimation theory,the maximum posterior estimation method is used to solve the tensor recovery problem.Under the alternating direction method of multipliers(ADMM)framework,tensor recovery is integrated to the video decoding process.In addition,to optimize the decoding algorithm,an adaptive parameter initialization algorithm is proposed.And the computation complexity of the proposed algorithm shows its efficiency.Finally,an improved version of the above algorithm is proposed.Specifically,the above algorithm is changed to a two-stage decoding algorithm.In the first stage,tensor recovery is used to protect video structure information,and in the second stage,matrix recovery is used to process smaller groups to optimize the details.Meanwhile,the performance verification of the two proposed algorithms is performed on both synthetic and real data.Experimental results show that the improved two-stage algorithm can achieve the state-of-the-art decoding results compared with the other algorithms.