Heterogeneous Face Recognition: A Common Encoding Feature Discriminant Approach.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biometrics

Background:

• Heterogeneous face recognition (HFR) involves matching face images across different modalities (e.g., infrared to visible light).
• Significant discrepancies between modalities pose a major challenge for conventional handcrafted feature descriptors.
• Existing methods often fail to extract common, discriminative information from heterogeneous face data.

Purpose of the Study:

• To propose a new feature descriptor, the common encoding model (CEM), for HFR.
• To reduce the modality gap at the feature extraction stage.
• To enhance cross-modality face recognition performance.

Main Methods:

• Developed a common encoding model to transform face images into an encoded representation.
• Learned the encoding model from training data to minimize differences between encoded heterogeneous faces of the same person.
• Proposed a discriminant matching method based on encoded face images for identity inference.

Main Results:

• The proposed CEM effectively captures common discriminant information across modalities.
• The approach significantly reduces the modality gap during feature extraction.
• Demonstrated effectiveness in NIR-to-VIS and sketch-to-photo face recognition scenarios on public datasets.

Conclusions:

• The common encoding model offers a robust solution for heterogeneous face recognition.
• The method successfully addresses the challenge of large modality discrepancies.
• The proposed approach enhances recognition performance in cross-modality face matching tasks.