Multi-Directional Multi-Level Dual-Cross Patterns for Robust Face Recognition.

Area of Science:

• Computer Vision
• Biometrics
• Pattern Recognition

Background:

• Unconstrained face recognition faces challenges due to variations in illumination, pose, and expression.
• Existing local descriptors often struggle to balance distinctiveness with robustness to these variations.

Purpose of the Study:

• To develop a novel feature extraction scheme for robust and accurate unconstrained face recognition.
• Introduce Multi-Directional Multi-Level Dual-Cross Patterns (MDML-DCPs) as an efficient and discriminative face descriptor.

Main Methods:

• Exploited the first derivative of Gaussian operator to mitigate illumination effects.
• Computed Dual-Cross Patterns (DCP) features at both holistic and component levels.
• Developed the MDML-DCPs scheme to encode invariant facial characteristics.

Main Results:

• MDML-DCPs demonstrated superior performance compared to state-of-the-art local descriptors (LBP, LTP, LPQ, etc.) on benchmark datasets (FERET, CAS-PERL-R1, FRGC 2.0, LFW).
• Achieved top performance in face identification and verification tasks, particularly on challenging datasets like LFW and FRGC 2.0.
• The DCP descriptor is computationally efficient, only doubling the cost of Local Binary Patterns computation.

Conclusions:

• MDML-DCPs offer a comprehensive and efficient method for encoding facial information, yielding high discriminative power and robustness.
• The proposed scheme provides a significant advancement for unconstrained face recognition systems.
• MDML-DCPs represent a promising approach for real-world face recognition applications.