Low-rank Representation For Video Signal Processing

In recent years,the low-rank representation method has been widely used for image and video processing,e.g.,low-rank representation-based video denoising and video foreground detection methods.However,many existing algorithms just directly apply the low-rank matrix recovery theory,and the simple models usually lead to poor performance.This thesis focusses on two typical video signal processing problems,for which two new low-rank representation-based models are respectively built.The achievements of our study are listed below:(1)In order to effectively remove the mix noise from video sequences,a new video denoising algorithm is proposed.This algorithm simultaneously exploits the local and nonlocal similarity among image blocks in a video sequence by utilizing total variation(TV)of the residual values and low-rank representation of groups of similar image blocks.First of all,block matching is applied in a noisy video sequence to find the most similar image blocks,after which similar image blocks are grouped together.Then,every group of similar image blocks is represented as the sum of a low-rank matrix and a sparse matrix.In addition,the TV minimization of residual values in the low-rank matrix is also required.Finally,the target optimization problem is efficiently solved so as to obtain the low-rank matrix,which is the final recovered group of image blocks.Experimental results show that the proposed algorithm can well remove both the Gaussian noise and the impulse noise.Compared with other algorithms,our method is able to achieve significantly higher peak signal-to-noise ratio(PSNR).(2)In order to accurately separate the foreground and background from a surveillance video clip,a new low rank and reweighted sparse decomposition model together with an optimization algorithm are proposed.In the proposed model,the foreground components are reweighted so that its sparsity can be enhanced.When establishing the weighting matrix,the optical flow method is used to get the motion vectors in each frame in order that the real moving areas can be recognized.When dealing with the optimization problem,the objective function is divided into several sub-problems,which are iteratively solved by using the split-Bregman technique.The experimental results show that the proposed algorithm can efficiently separate foreground and background components for video clips with or without noises.When compared with other existing algorithms,the proposed algorithm achieves significantly better objective and subjective results.(3)Based on the proposed algorithm,we develop a low-rank representation-based video denoising system and a low-rank representation-based video foreground detection system on Matlab platform.The two developed systems have simple human-machine interfaces,and can well display the performance of our proposed algorithms.To further evaluate the proposed algorithms,the users can adjust the parameters according to their requirements.