Robust Cross-Domain Pseudo-Labeling and Contrastive Learning for Unsupervised Domain Adaptation NIR-VIS Face...

Area of Science:

• Computer Science
• Artificial Intelligence
• Biometrics

Background:

• Near-infrared and visible (NIR-VIS) face recognition is crucial for 24-hour security, especially in low-light conditions.
• Manual annotation of large, heterogeneous face datasets for NIR-VIS recognition is costly and time-consuming, hindering real-world applications.
• Unsupervised domain adaptation offers a promising approach to overcome annotation limitations in NIR-VIS face recognition.

Purpose of the Study:

• To develop an unsupervised domain adaptation method for NIR-VIS face recognition.
• To eliminate the need for manual identity labels in large-scale NIR-VIS face recognition systems.
• To propose a novel network architecture that effectively handles domain discrepancies between NIR and VIS face images.

Main Methods:

• A novel Robust cross-domain Pseudo-labeling and Contrastive learning (RPC) network is proposed.
• Key components include NIR cluster-based Pseudo labels Sharing (NPS) for generating reliable pseudo-labels, Domain-specific cluster Contrastive Learning (DCL) for learning discriminative intra-domain representations, and Inter-domain cluster Contrastive Learning (ICL) for robust, domain-independent feature learning.
• The NPS component leverages NIR clusters to share label knowledge with the VIS domain, while DCL and ICL refine representations through contrastive learning strategies.

Main Results:

• The proposed RPC network achieved over 99% accuracy in pseudo-label assignment.
• Experimental results on four mainstream NIR-VIS datasets demonstrate the advanced performance of the RPC network.
• The method successfully learns robust and domain-independent representations without manual annotations.

Conclusions:

• The RPC network effectively addresses the challenges of unsupervised domain adaptation in NIR-VIS face recognition.
• The proposed approach significantly reduces the reliance on manual annotations, paving the way for scalable real-world applications.
• This work advances the state-of-the-art in NIR-VIS face recognition by enabling accurate and efficient recognition under varying lighting conditions.