A NIR-to-VIS face recognition via part adaptive and relation attention module

TL;DR

This paper introduces a novel part relation attention module and adaptive triplet loss for NIR-to-VIS face recognition, effectively handling domain gaps, pose, and emotion variations, and demonstrating superior performance on benchmark datasets.

Contribution

It proposes a new part relation attention module and adaptive triplet loss to improve heterogeneous face recognition across domains, pose, and emotion variations.

Findings

Improved accuracy on CASIA NIR-VIS 2.0 dataset.

Achieved state-of-the-art results on BUAA-VisNir dataset.

Effectively handles pose and emotion variations.

Abstract

In the face recognition application scenario, we need to process facial images captured in various conditions, such as at night by near-infrared (NIR) surveillance cameras. The illumination difference between NIR and visible-light (VIS) causes a domain gap between facial images, and the variations in pose and emotion also make facial matching more difficult. Heterogeneous face recognition (HFR) has difficulties in domain discrepancy, and many studies have focused on extracting domain-invariant features, such as facial part relational information. However, when pose variation occurs, the facial component position changes, and a different part relation is extracted. In this paper, we propose a part relation attention module that crops facial parts obtained through a semantic mask and performs relational modeling using each of these representative features. Furthermore, we suggest component adaptive triplet loss function using adaptive weights for each part to reduce the intra-class identity regardless of the domain as well as pose. Finally, our method exhibits a performance improvement in the CASIA NIR-VIS 2.0 and achieves superior result in the BUAA-VisNir with large pose and emotion variations.