Face Recognition via Collaborative Representation: Its Discriminant Nature and Superposed Representation.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Collaborative representation methods like Sparse Subspace Clustering (SSC) and Sparse Representation-based Classification (SRC) excel in face recognition on controlled datasets.
• Performance degrades on uncontrolled or undersampled datasets due to misleading coefficients from incorrect classes.

Purpose of the Study:

• To address the limitations of collaborative representation on challenging datasets.
• To develop a novel classifier that improves generalization ability in face recognition.

Main Methods:

• Inspired by Linear Discriminant Analysis (LDA), a Superposed Linear Representation Classifier (SLRC) was developed.
• SLRC represents test images as a superposition of class centroids and shared intra-class differences.
• Enforced sparsity constraint for enhanced performance.

Main Results:

• SLRC significantly improves the generalization ability of collaborative representation.
• SLRC demonstrates competitive performance against advanced dictionary learning techniques on AR and FRGC databases.
• Achieved state-of-the-art results on the FERET database with single sample per person when enforced with sparsity.

Conclusions:

• SLRC offers a robust solution for face recognition challenges posed by uncontrolled and undersampled data.
• The proposed method enhances the effectiveness of collaborative representation, particularly in low-data regimes.
• SLRC provides a strong baseline and achieves top-tier performance, especially with sparsity constraints.