Weighted Mixed-Norm Regularized Regression for Robust Face Identification.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Regression-based classification is key for face identification (FI).
• Existing methods excel with either noncontiguous (vector-based) or contiguous (matrix-based) noise.
• A gap exists in handling mixed contiguous and noncontiguous noise.

Purpose of the Study:

• To propose a novel method for face identification under mixed noise corruption.
• To bridge the gap between vector-based and matrix-based regression methods.
• To enhance both accuracy and computational efficiency in face identification.

Main Methods:

• Developed Weighted Mixed-Norm Regression (WMNR) to address mixed image noise.
• Unified residual distributions with a feature-weighted loss function.
• Constrained residual images using low-rank structures and a weight factor.
• Derived a reweighted alternating direction method of multipliers algorithm for efficient computation.

Main Results:

• WMNR demonstrates superior performance over state-of-the-art regression-based methods.
• Achieved over 10% performance improvement and >70% runtime savings compared to vector-based methods.
• Obtained up to 20% performance improvement compared to matrix-based methods, despite slightly higher computation.

Conclusions:

• WMNR effectively handles mixed noise in face identification.
• Offers a significant improvement in identification accuracy and computational efficiency.
• Represents a promising advancement for robust face recognition systems.