Sparse and Dense Hybrid Representation via Dictionary Decomposition for Face Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Sparse representation-based classification (SRC) is effective when training data is sufficient and uncorrupted.
• Practical applications like face identification often lack sufficient or uncorrupted training samples per class.
• Violations of these conditions degrade SRC performance significantly.

Purpose of the Study:

• To address the limitations of SRC when training data is insufficient or corrupted.
• To propose a novel framework that enhances classification robustness.
• To improve the performance of face identification systems under challenging conditions.

Main Methods:

• Developed a sparse- and dense-hybrid representation (SDR) framework.
• Introduced a supervised low-rank (SLR) dictionary decomposition procedure.
• Integrated SDR with SLR for enhanced classification and corrupted data handling.

Main Results:

• The proposed SDR-SLR approach significantly improves classification performance.
• Effectiveness demonstrated in face recognition applications.
• Outperforms existing state-of-the-art sparse representation methods on benchmark datasets.

Conclusions:

• The SDR-SLR framework effectively alleviates problems associated with sparse representation.
• The approach offers advancements in face recognition accuracy and robustness.
• Validated through extensive experiments on benchmark face databases.