Coupled Attribute Learning for Heterogeneous Face Recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biometrics

Background:

• Heterogeneous face recognition (HFR) faces challenges due to significant textural and spatial differences in face images from varied sources.
• Existing HFR methods primarily focus on cross-modality image matching, overlooking the practical availability of semantic facial attribute descriptions.
• Limited training data for cross-modality pairs further complicates HFR due to complex image generation processes.

Purpose of the Study:

• To propose a novel coupled attribute learning for HFR (CAL-HFR) method that leverages facial attributes to address modality gaps.
• To develop an end-to-end HFR network that learns attribute-identity relationships without manual attribute labeling.
• To improve the robustness and accuracy of face recognition across different imaging modalities.

Main Methods:

• Utilizing deep convolutional networks to map heterogeneous face images into a shared compact feature space.
• Introducing a coupled attribute guided triplet loss (CAGTL) to train the network, effectively correcting attribute estimation errors.
• Employing inherent relationships between face attributes and identities for enhanced recognition.

Main Results:

• The proposed CAL-HFR method demonstrates superior performance over state-of-the-art methods in various heterogeneous scenarios.
• The network successfully learns to utilize facial attributes for improved cross-modality face matching.
• The method effectively handles defects arising from incorrectly estimated attributes.

Conclusions:

• CAL-HFR offers a significant advancement in heterogeneous face recognition by integrating attribute learning.
• The approach effectively bridges the gap between different face image modalities using semantic attribute information.
• A publicly available annotated heterogeneous facial attribute database is released to support future research.