DVG-Face: Dual Variational Generation for Heterogeneous Face Recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Heterogeneous face recognition (HFR) is vital for security but faces challenges due to domain discrepancies and limited data.
• Existing methods struggle with the scarcity of paired heterogeneous face images for training.

Purpose of the Study:

• To address data scarcity and domain gaps in HFR.
• To develop a novel framework for generating diverse, identity-consistent heterogeneous face image pairs.
• To improve the performance of HFR systems through enhanced training data and contrastive learning.

Main Methods:

• Formulating HFR as a dual generation problem using a novel dual variational generation (DVG-Face) framework.
• Designing a dual variational generator to learn the joint distribution of paired heterogeneous images.
• Integrating identity information from large-scale visible data and employing a pairwise identity preserving loss.
• Utilizing generated diverse paired heterogeneous images for training HFR networks via contrastive learning.

Main Results:

• DVG-Face successfully generates massive, diverse paired heterogeneous images with consistent identities from noise.
• The generated images enable training HFR networks that yield domain-invariant and discriminative embedding features.
• The proposed method achieves superior performance across seven challenging databases in five HFR tasks (NIR-VIS, Sketch-Photo, Profile-Frontal Photo, Thermal-VIS, ID-Camera).

Conclusions:

• DVG-Face effectively overcomes limitations of small-scale paired heterogeneous data in HFR.
• The framework's ability to generate diverse and consistent identity pairs significantly enhances HFR model training.
• This approach offers a robust solution for improving HFR accuracy in real-world security applications.