A Truncated Nuclear Norm Regularization Method Based on Weighted Residual Error for Matrix Completion.

Area of Science:

• Computer Vision
• Matrix Analysis
• Optimization

Background:

• Low-rank matrix completion is crucial in computer vision for recovering incomplete matrices.
• Existing nuclear norm minimization methods face limitations in approximating matrix rank.
• Truncated nuclear norm regularization (TNNR) offers improved rank approximation but suffers from sensitivity and slow convergence.

Purpose of the Study:

• To develop novel TNNR-based methods for enhanced matrix completion.
• To address the convergence speed and robustness issues of existing TNNR methods.
• To improve the accuracy and efficiency of recovering low-rank matrices from limited data.

Main Methods:

• Proposed TNNR with Weighted Residual Error (TNNR-WRE) using an augmented Lagrange function with weighted residual errors.
• Developed an Extended TNNR-WRE (ETNNR-WRE) model for increased robustness.
• Compared performance against TNNR, TNNR alternating direction method of multipliers, TNNR accelerated proximal gradient with Line search, and Iteratively Reweighted Nuclear Norm (IRNN).

Main Results:

• TNNR-WRE accelerates convergence compared to standard TNNR.
• ETNNR-WRE demonstrates superior robustness to the number of subtracted singular values.
• Both TNNR-WRE and ETNNR-WRE outperform TNNR and IRNN on synthetic and real visual datasets.

Conclusions:

• TNNR-WRE and ETNNR-WRE offer significant improvements in matrix completion efficiency and reliability.
• The proposed methods provide effective solutions for low-rank matrix recovery in computer vision.
• ETNNR-WRE is particularly robust, making it suitable for diverse visual data applications.