视频背景分离中一种新的非凸秩近似的RPCA模型

A new non-convex rank approximation RPCA model for video background separation

视频背景分离以及前景提取广泛应用于场景分析、目标追踪等领域。而鲁棒主成分分析(RPCA)则是实现视频背景与前景分离的重要技术之一。但是,用核范数来近似秩函数的传统RPCA模型在处理含有较大奇异值的图像时效果并不理想。为了解决这一问题,提出一种新的非凸函数来近似秩函数,对基于核范数的RPCA模型进行改进,并应用增广拉格朗日乘子法求解改进的模型。实验结果表明,与传统的RPCA及现有的一些改进模型相比,提出的基于非凸秩近似的RPCA模型计算效率更高,且图像分离效果更好。

Video background separation and foreground extraction are widely used in scene analysis,target tracking and other fields.Robust principal component analysis(RPCA)is one of the important techniques to realize the separation of video background and foreground.However,the traditional RPCA model that approximates the rank function by the nuclear norm is not effective in dealing with images with large singular values.In order to overcome this problem,this paper proposed a new non-convex function to approximate the rank function to improve the RPCA model based on nuclear norm.The extended Lagrange multiplier method was used to solve the improved RPCA model.Experimental results show that compared with the traditional RPCA and few other improved models,the non-convex approximation model proposed in this paper has higher computation efficiency and better image separation effect.