Half-Face Dictionary Integration for Representation-Based Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Pattern Recognition

Background:

• Representation-based classification is crucial for facial recognition.
• Existing methods face challenges with limited training data and high dimensionality.
• The small sample size problem significantly impacts classification accuracy.

Purpose of the Study:

• To propose a novel Half-Face Dictionary Integration (HFDI) algorithm.
• To enhance the accuracy and robustness of face classification, especially under small sample size conditions.
• To improve the collaborative representation of test samples using augmented training data.

Main Methods:

• Generation of virtual half-face training samples for data augmentation.
• Development of a dual-column (row) half-face integrated dictionary.
• Implementation of an elimination strategy to select robust training atoms and reduce dictionary redundancy.
• Application of a competitive fusion method to weight reconstruction residuals for classification.

Main Results:

• The HFDI algorithm demonstrated effectiveness across multiple benchmark datasets (FRT, Aleix and Robert, Georgia Tech, ORL, CMU-PIE).
• Significant improvements in face classification accuracy were observed, particularly when dealing with limited training samples.
• The method successfully addressed the challenges posed by the small sample size problem in face recognition.

Conclusions:

• The proposed HFDI algorithm offers a robust and effective solution for representation-based face classification.
• Data augmentation via virtual half-face samples and dictionary integration are key to improving performance.
• HFDI shows strong potential for real-world facial recognition applications, especially in scenarios with scarce training data.