Research On Low-Rank And Sparse Decomposition Based On Weighted Norm And Mixed Sparse Prior

Computer vision,as one of the three main research areas of artificial intelligence,has attracted extensive attention and has become its research hotspot.Video foreground-background separation has become one of the core problems which need to be solved imperatively in computer vision.In computer vision,the excellent performance of low-rank and sparse models makes it become a powerful tool.To tackle the problems of video foreground-background separation,this paper has completed mainly the following work:Firstly,three low-rank and sparse decomposition algorithms based on the Laplace distribution have been proposed for video foreground-background separation.By introducing Laplacian scale mixture modeling of sparse foreground,the sparse balance parameters are adaptively adjusted,thereby improving the adaptability of the low-rank and sparse decomposition algorithm.More foreground information can be recovered in many scenes while reducing the difficulty of model tuning.Simulation results show that compared with mainstream low-rank and sparse decomposition algorithms,our proposed algorithm can recover more complete foregrounds in classical scenes and is closer to the real foreground condition.The average values of its are 0.065,0.06,and 0.081 higher than the comparative algorithms of the same type,respectively.Secondly,a new low-rank and sparse decomposition model based on the spectral norm and structural sparsity-inducing norm with stronger robustness is proposed and applied to complex scenarios with video foreground-background separation.The proposed algorithm utilizes the spectral norm to strengthen the low-rank constraint on the static background and uses the structural sparsity-inducing norm to describe the sparse item.The sparse item is further decomposed into a dynamic background item and a real foreground item,and the norm and the total variation regularized are used for modeling respectively,to obtain a clean foreground image under complex background.Simulation experiments show that the of the proposed algorithm attained five of the highest and three of the second highest values in eight noise-free test video sequences with complex backgroundscan,while the highest average were also achieved in all seven experimental groups with noise.In conclusion,given the problems existing in video foreground-background separation,the proposed algorithm can recover more detailed foreground information compared with mainstream algorithms and has wider adaptability and stronger robustness.