Modal Regression-Based Atomic Representation for Robust Face Recognition and Reconstruction.

Area of Science:

• Computer Science
• Artificial Intelligence
• Pattern Recognition

Background:

• Representation-based classification (RC) methods, including sparse RC, show promise for face recognition (FR).
• Conventional RC methods rely on regression models with predefined noise distribution assumptions (e.g., Gaussian, Laplacian).
• These assumptions are often violated by complex real-world noise, limiting performance.

Purpose of the Study:

• To propose a robust modal regression (MR)-based atomic representation and classification (MRARC) framework.
• To overcome limitations of existing RC methods caused by violated noise distribution assumptions.
• To enhance face recognition performance in the presence of complex noise.

Main Methods:

• Developed a modal regression (MR)-based atomic representation and classification (MRARC) framework.
• MRARC regresses toward the conditional mode function, offering robustness without specific noise distribution assumptions.
• Utilized atomic representation, a general framework encompassing sparse, collaborative, and low-rank representations.
• Devised a general optimization algorithm using alternating direction method of multipliers and half-quadratic theory.

Main Results:

• The MRARC framework demonstrates capability in handling various complex noises.
• Four novel RC methods for unimodal and multimodal FR were developed using the MRARC platform.
• Experiments on real-world data validated the MRARC framework's efficacy.
• MRARC achieved robust face recognition and reconstruction.

Conclusions:

• The proposed MRARC framework provides a robust approach to face recognition.
• MRARC effectively addresses limitations of traditional RC methods by not assuming specific noise distributions.
• The framework offers a versatile platform for developing advanced representation-based classification methods.