基于截断核范数的视频前景与背景分离

Video foreground-background separation based on truncated nuclear norm

为解决鲁棒主成分分析模型(RPCA)中核范数不是矩阵秩函数最佳近似的问题,提出截断核鲁棒主成分分析模型(TNSRPCA)。使用截断核范数替代传统的核范数进行低秩约束,将稀疏部分添加正则化约束克服动态背景产生的扰动,为解决截断核范数使用两步迭代法计算量巨大的问题,提出部分奇异值阈值算子进行求解,提高计算效率。实验结果表明,该模型在视频前景与背景分离中获得了较好的分离效果和抗噪声鲁棒性。

To deal with the problem that the nuclear norm of the robust principal component analysis model（RPCA）is not the best approximation of the matrix rank function,a model called truncated nuclear robust principal component analysis model（TNSRPCA）was proposed.The nuclear norm was replaced by truncated nuclear norm for low rank constraint.The regular constraint for sparse was added to remove the dynamic background interference.The approach consistently uses a two-step iterative strategy to recover the low rank components,which needs huge computation work.To solve this problem,apartial singular value thresholding operator was proposed.As a result,computing efficiency was improved.Experimental results show that the proposed models achieve better effects on the foreground-background separation,and possess superior robustness against noise.