Heterogeneous Face Recognition by Margin-Based Cross-Modality Metric Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Heterogeneous face recognition (HFR) involves matching face images from different sources or modalities.
• Key challenges in HFR include cross-modal variations and achieving subject separation across modalities.

Purpose of the Study:

• To propose a novel margin-based cross-modality metric learning (MCM2L) method for HFR.
• To develop a common subspace for measuring cross-modality distances.
• To enhance the separation between intra-personal and inter-personal distances across modalities.

Main Methods:

• Defined a cross-modality metric in a shared subspace.
• Implemented two constraints: minimizing intra-personal cross-modality distances and enforcing a margin between intra-personal and inter-personal distances.
• Utilized hinge loss on triplet-based constraints for optimization, focusing on difficult-to-separate cases.
• Extended the method to a kernelized version (KMCM2L).

Main Results:

• Evaluated MCM2L and KMCM2L on ID card and benchmark cross-modality face datasets.
• Incorporated various feature extraction methods, including deep learned features.
• Demonstrated marked improvements over state-of-the-art methods in most experimental settings.

Conclusions:

• MCM2L and KMCM2L effectively address the challenges of heterogeneous face recognition.
• The proposed metric learning approaches significantly enhance cross-modality face matching performance.
• The methods show robustness across different feature extraction techniques.