Enhanced Group Sparse Regularized Nonconvex Regression for Face Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Regression-based methods are successful in face recognition but may be suboptimal due to convex relaxations (e.g., l1-norm, nuclear norm).
• These relaxations can introduce bias and are sensitive to outliers, limiting robustness.

Purpose of the Study:

• To propose a novel Enhanced Group Sparse regularized Nonconvex Regression (EGSNR) method for robust face recognition.
• To address the limitations of convex relaxations and improve recognition accuracy in the presence of complex errors and outliers.

Main Methods:

• Introduced an upper bounded nonconvex function to replace the l1-norm for sparsity, mitigating bias and outlier effects.
• Developed a mixed model using gamma-norm and matrix gamma-norm to capture complex error characteristics.
• Designed an l2,γ-norm regularizer for interclass or group sparsity, enhancing discriminative representation coefficients.
• Incorporated data locality by considering the distance between query samples and multi-subspaces.

Main Results:

• The proposed EGSNR method demonstrated superior performance compared to state-of-the-art regression-based face recognition techniques.
• Experiments on multiple benchmark face datasets validated the effectiveness of the enhanced group sparse regularizer.
• The nonconvex approach effectively alleviated bias and reduced the adverse impact of outliers.

Conclusions:

• EGSNR offers a more robust and accurate approach to face recognition by overcoming limitations of traditional convex regression methods.
• The novel regularization techniques and nonconvex formulation contribute to learning more discriminative features for improved recognition.
• The method shows significant potential for real-world applications requiring reliable face recognition under challenging conditions.