Learning discriminant face descriptor.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Local feature descriptors like Gabor and Local Binary Patterns (LBP) are crucial for face recognition but traditionally handcrafted.
• Existing methods often rely on predefined descriptor forms, limiting adaptability and performance.

Purpose of the Study:

• To propose a data-driven method for learning a Discriminant Face Descriptor (DFD).
• To enhance face representation discriminability by learning filters, optimizing sampling, and constructing dominant patterns.
• To adapt the DFD for heterogeneous (cross-modality) face recognition by developing a coupled DFD (C-DFD).

Main Methods:

• Learning discriminant image filters to enhance feature extraction.
• Soft determination of optimal neighborhood sampling strategies.
• Statistical construction of dominant patterns for robust feature representation.
• Incorporating discriminative learning for effective and robust feature extraction.
• Developing coupled DFD (C-DFD) to bridge the feature gap in heterogeneous face recognition.

Main Results:

• The proposed DFD method significantly improves face recognition performance.
• DFD enhanced POEM and LQP accuracy by approximately 4.5% on the LFW database.
• C-DFD improved heterogeneous face recognition performance over LBP by over 25%.
• Experiments validated effectiveness on FERET, CAS-PEAL-R1, LFW, and HFB databases.

Conclusions:

• The data-driven approach to learning DFD is effective for both homogeneous and heterogeneous face recognition.
• Learned discriminant filters, optimized sampling, and dominant pattern construction enhance feature robustness.
• C-DFD effectively addresses the challenges of cross-modality face recognition, reducing the feature domain gap.