Learning Compact Binary Face Descriptor for Face Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Biometrics

Background:

• Binary feature descriptors like Local Binary Patterns (LBP) are vital for face recognition due to robustness and discriminative power.
• Existing descriptors are often hand-crafted, necessitating significant domain expertise.
• There is a need for automated, efficient binary feature learning methods for face representation.

Purpose of the Study:

• To propose a novel compact binary face descriptor (CBFD) learning method for enhanced face representation and recognition.
• To develop an unsupervised feature learning approach that generates discriminative and compact binary codes.
• To introduce a coupled CBFD (C-CBFD) for addressing the modality gap in heterogeneous face recognition.

Main Methods:

• Extracting pixel difference vectors (PDVs) from local image patches.
• Unsupervised learning of a feature mapping to project PDVs into low-dimensional binary codes, optimizing variance, minimizing reconstruction loss, and ensuring even code distribution.
• Clustering and pooling binary codes into histogram features for final face representation.
• Developing C-CBFD to reduce feature-level modality gaps for heterogeneous recognition.

Main Results:

• The proposed CBFD method effectively learns compact and discriminative binary codes.
• CBFD demonstrates superior performance compared to traditional hand-crafted binary descriptors.
• The C-CBFD variant shows significant improvements in heterogeneous face recognition tasks.
• Experiments on five benchmark datasets validate the effectiveness of both CBFD and C-CBFD.

Conclusions:

• The proposed CBFD feature learning method offers a powerful alternative to hand-crafted descriptors for face recognition.
• The unsupervised learning approach effectively captures essential facial information in compact binary codes.
• C-CBFD provides a viable solution for heterogeneous face recognition by bridging modality gaps.
• The methods show state-of-the-art performance, highlighting their potential for real-world applications.