Locality-Constrained Discriminative Matrix Regression for Robust Face Identification.

Area of Science:

• Computer Science
• Machine Learning
• Biometrics

Background:

• Regression-based methods are common in face identification, approximating query samples via linear combinations of training data.
• Existing matrix regression models offer robustness to noise but overlook crucial label information and local data relationships.

Purpose of the Study:

• To propose a novel robust representation method, locality-constrained discriminative matrix regression (LDMR), for enhanced face identification.
• To address limitations of prior methods by integrating label information and local data structure.

Main Methods:

• LDMR directly constrains representation components, considering label information for a closer link to the identification process.
• It utilizes subspace distances to characterize locality structure for learning class weights, prioritizing correct class contributions.
• Class weights are integrated into a competitive constraint to reduce inter-class correlations and strengthen intra-class relationships.

Main Results:

• LDMR demonstrates superior performance compared to existing state-of-the-art regression-based methods.
• Experiments on benchmark datasets validate the effectiveness of the proposed LDMR approach.

Conclusions:

• LDMR offers a robust and effective approach to face identification by leveraging label information and local data structure.
• The method shows significant improvements over previous regression-based techniques in experimental evaluations.