Robust Low-Rank Matrix Factorization Under General Mixture Noise Distributions.

Area of Science:

• Computer Vision
• Machine Learning
• Data Science

Background:

• Low-rank matrix factorization (LRMF) is a key technique for subspace learning in high-dimensional data.
• Existing LRMF methods primarily address Gaussian or Laplace noise using L2 or L1 norms.
• There is a need for LRMF models that can adapt to more complex and varied noise distributions.

Purpose of the Study:

• To develop a more robust LRMF model capable of handling complex noise patterns.
• To introduce a novel LRMF framework that automatically fits real-world noise characteristics.
• To enhance LRMF performance in computer vision applications by addressing noise limitations.

Main Methods:

• Proposed a new LRMF model based on mixture of exponential power (MoEP) distributions to capture complex noise.
• Introduced a penalized MoEP (PMoEP) model by integrating penalized likelihood with MoEP distributions.
• Developed the PMoEP-Markov Random Field (PMoEP-MRF) model by incorporating local noise continuity using MRFs.
• Designed Generalized Expectation Maximization (GEM) and variational GEM algorithms for parameter inference.

Main Results:

• The proposed MoEP-based LRMF models demonstrate superior performance in handling complex noise compared to traditional methods.
• Extensive experiments validated the effectiveness on synthetic data, face modeling, hyperspectral image denoising, and background subtraction.
• The models showed an ability to automatically adapt to diverse noise types through the MoEP distribution fitting.

Conclusions:

• The novel PMoEP and PMoEP-MRF models offer significant advancements in LRMF for computer vision by effectively modeling complex noise.
• These methods provide a more flexible and robust approach to subspace learning in the presence of non-standard noise.
• The findings suggest broader applicability of these techniques in various image processing and computer vision tasks.