Research About Video Reconstruction Methods Based On Low-Rank Approximation And Dictionary Learning

Nowadays,big data suffers rapid growth in data volume,over half of which is video data.In order to get reliable data efficiently,the requirements for the stability of transforming channel and the high equality of the recovery algorithm have became increasingly demanded.The traditional video compression methods are based on Shannon's theorem,so that the collected signal can be only recovered if its collected frequency is higher than the Nyquist sampling rate.The emerging compressive sensing theory provides a new signal processing method,which breaks through the limitation of Shannon's theory and achieves the acquisition and restoration of data at a frequency lower than the Nyquist sampling rate.In this paper,based on compressive sensing,using dictionary learning method as a framework,combining the multi-dimensional similarity of video data,the research focuses on the video compression and reconstruction method in the coded aperture temporal compressive imaging system.The main contents are as follows:Firstly,the basic knowledge of compressive sensing and the sparse representation methods based on dictionary are introduced.The applications of compressive sensing in the field of video reconstruction are described.The mathematical model of temporal coded aperture compressive imaging system is analyzed,and its superiority in sampling rate and system cost,and its diversity in calculation method are pointed out.Secondly,a dual resolution sparse dictionary learning method is proposed.First,the low-resolution copy of the dictionary training data of a specific scene can be got by interpolation.Then,combined with the original relatively high-resolution training data,the dual resolution dictionary can be learned through the restriction of keeping the two corresponding sparse representations the same.Sparse representation of the restoration data can be got with the low-resolution dictionary,then combine it with the high-resolution dictionary,so that the reconstruction can be done.The dual resolution dictionary learning method can be applied in both two-dimension image denoising and three-dimension video reconstruction,and experimental results show that the method can guarantee stable and accurate signal reconstruction in both applications.Finally,video data has a natural multi-dimensional structure,and the images of each frame have non-local similarity and the continuous sequences have strong spatial-temporal correlation.According to the structural characteristics of video data,this paper fully exploits the multi-dimensional sparse characteristics of data using compressive sensing,and proposes a signal reconstruction method for multi-dimensional video data.This method designs an online video frame image block clustering method,which constructs a low rank matrix and conducts low-rank approximation.Combining compressive sensing and dictionary learning,video data can be reconstructed.The simulation results demonstrate that the proposed method has higher quality reconstruction performance on the changing scene while guaranteeing high compression rate.