Double Nonconvex Tensor Robust Kernel Principal Component Analysis and Its Visual Applications.

Area of Science:

• Computer Vision
• Machine Learning
• Data Science

Background:

• Tensor robust principal component analysis (TRPCA) is a linear method for visual tasks.
• TRPCA assumes low-rankness, which is often violated by nonlinear tensor data, leading to approximation errors.
• Nonlinear structures in tensor data require advanced decomposition methods beyond linear assumptions.

Purpose of the Study:

• To establish a general paradigm for nonlinear tensor decomposition, termed tensor robust kernel principal component analysis (TRKPCA).
• To develop novel nonconvex regularizers, kernelized tensor Schatten-p norm (KTSPN) and generalized nonconvex regularization, for TRKPCA.
• To propose a double nonconvex TRKPCA (DNTRKPCA) method integrating these regularizers for enhanced performance.

Main Methods:

• Development of TRKPCA for nonlinear tensor data.
• Introduction of KTSPN to capture implicit low-rankness and nonlinear features.
• Design of generalized nonconvex regularization for sparser structural coding.
• Implementation of DNTRKPCA using an alternating direction multiplier method (ADMM) optimization framework.

Main Results:

• The proposed DNTRKPCA method effectively captures nonlinear features and achieves robust separation.
• Experimental results demonstrate the superior performance of DNTRKPCA compared to state-of-the-art regularization methods.
• The method shows high competitiveness on both synthetic and real-world datasets.

Conclusions:

• DNTRKPCA offers a more effective approach for analyzing nonlinear tensor data than traditional TRPCA.
• The novel nonconvex regularizers significantly improve the robustness and accuracy of tensor decomposition.
• The developed ADMM framework provides an efficient solution for the proposed TRKPCA method.