Fast and robust rank-one matrix completion via maximum correntropy criterion and half-quadratic optimization

Zhi-Yong Wang*, Hing Cheung So, Zhaofeng Liu

*Corresponding author for this work

Research output: Journal Publications and ReviewsRGC 21 - Publication in refereed journalpeer-review

10 Citations (Scopus)

Abstract

Matrix completion refers to seeking a low-rank matrix to match an incomplete matrix and fill its missing entries, and is an important topic because many real-world data can be modeled as low-rank matrices. Most existing schemes rely on ℓ2-norm minimization and hence are not robust to outliers. Even the ℓ1-norm based robust matrix completion algorithms proposed in the literature have poor recovery performance for large gross errors, and require knowing the matrix rank, which is difficult to accurately determine in practice. In this paper, we devise a robust and fast rank-one matrix completion algorithm via combining the maximum correntropy criterion (MCC) and half-quadratic (HQ) optimization theory. The MCC, i.e., minimizing the Welsch cost function, can resist the gross errors but it is non-convex. While HQ optimization can transform the Welsch cost function into a quadratic form, making the resultant optimization problem easy to solve. Furthermore, an adaptive kernel width strategy is derived and there are no tunable parameters except for the termination conditions in our algorithm. Computer simulations using synthetic data and experimental results of real-world images demonstrate that the developed method achieves accurate recovery performance and high computational efficiency.
Original languageEnglish
Article number108580
JournalSignal Processing
Volume198
Online published13 Apr 2022
DOIs
Publication statusPublished - Sept 2022

Research Keywords

  • Adaptive kernel
  • Half-quadratic optimization
  • Maximum correntropy
  • Rank-one
  • Robust matrix completion

Fingerprint

Dive into the research topics of 'Fast and robust rank-one matrix completion via maximum correntropy criterion and half-quadratic optimization'. Together they form a unique fingerprint.

Cite this