A Unified View of Robust Decomposition in Low-rank + Additive Matrices
Low-rank and sparse matrix decomposition from noisy matrix observations (Image from Lu et al. 2019)
1. Introduction
Problem formulations of robust subspace learning/tracking by decomposition into low-rank plus additive matrices show a suitable framework for image and video analysis. The representative problem formulations is the Robust Principal Component Analysis (RPCA) solved via Principal Component Pursuit which decomposes a data matrix in a low-rank matrix and sparse matrix. However, similar implicit or explicit decompositions can be achieved in the other following problem formulations: Robust Non-negative Matrix Factorization (RNMF), Robust Subspace Recovery (RSR), Subspace Tracking (ST) , Robust Matrix Completion (RMC) and Robust Low-Rank Minimization (RLRM).
2. Robust Decomposition in Low-rank + Additive Matrice: Principle
All the decompositions in these different problem formulations of robust subspace learning/tracking can be considered in a unified view that we called Decomposition into Low-rank plus Additive Matrices (DLAM). Thus, all the decompositions can be written in a general formulation as follows:
The first matrix M1 is a low-rank matrix L.
The second matrix M2 is an unconstrained (residual) matrix S in RLRM and RMC, and a sparse matrix S in RPCA, RNMF, RSR and ST.
The third matrix M3 is generally the noise matrix E. The noise can be modeled by a Gaussian, a mixture of Gaussians (MoG) or a Laplacian distribution. E can capture turbulences in the case of background/foreground separation .
Practically, the decomposition is implicit when K=1. It is the degenerated case for problem formulations in their basic formulation (LRM, MC, etc...) which are not robust because there are no constraints on the matrix S=A-L.
The decomposition is explicit when K =2 and we have A=L+S. It is the for problem formulations in their robust formulation (RLRM, RMC, RPCA, RNMF, RSR, RST).
In the case of K=3, the decomposition is A = L+S +E. This explicit decomposition is called ”stable” decomposition as it separates the outliers in S and the noise in E.
Figure 1 shows an overview of the proposed unified view.
Figure 1: A Unified View of Problem Formulations via Decomposition into Low-rank plus Additive Matrices (DLAM)
3. Key Characteristics
4. Applications
The different robust problem formulations based on the decomposition into low-rank plus additive matrices are fundamental in several applications . Indeed, as this decomposition is nonparametric and does not make many assumptions, it is widely applicable to a large scale of problems ranging from:
4.1 Applications in Statistics
4.2 Applications in Signal Processing (one-dimensional signal)
Communication (Satellite)
Seismology (wave separation, wave denoising)
Radars (Moving Target Indication, Buried Object Detection, Interference Suppression)
4.3 Applications in Computer Vision (two-dimensional signal) (three-dimensional signal)
Image processing (image denoising, image composition, image colorization, image alignment and rectification, multi-focus image, face recognition)
Video processing (action recognition, motion estimation, motion saliency detection, video coding, key frame extraction, hyperspectral video processing, video restoration, background and foreground separation)
3D computer vision (Structure from Motion, 3D motion recovery, 3D reconstruction)
4.4 Applications in Computer Graphics
Rendering
4.5 Applications in Computer Science
4.6 Applications in Astronomy
4.7 Applications in Industrial Process
Author: Thierry BOUWMANS, Associate Professor, Lab. MIA, Univ. Rochelle, France.
Fair Use Policy
As this website gives many information that come from my research, please cite my following survey papers:
T. Bouwmans, A. Sobral, S. Javed, S. Jung, E. Zahzah, "Decomposition into Low-rank plus Additive Matrices for Background/Foreground Separation: A Review for a Comparative Evaluation with a Large-Scale Dataset", Computer Science Review, Volume 23, pages 1-71, February 2017. [pdf]
T. Bouwmans, E. Zahzah, “Robust PCA via Principal Component Pursuit: A Review for a Comparative Evaluation in Video Surveillance”, Special Issue on Background Models Challenge, Computer Vision and Image Understanding, CVIU 2014, Volume 122, pages 22–34, May 2014. [pdf]
Note: My publications are available on Academia, ResearchGate, Researchr, ORCID and Publication List.