Heterogeneous face recognition using kernel prototype similarities.

Area of Science:

• Computer Science
• Biometrics
• Artificial Intelligence

Background:

• Heterogeneous face recognition (HFR) matches face images from different modalities (e.g., infrared to photograph).
• Accurate HFR is crucial for forensics and surveillance, where gallery images are often photographs and probe images are from alternate sources.
• Existing HFR systems face challenges with varying image modalities and small sample sizes.

Purpose of the Study:

• To propose a generic and accurate framework for heterogeneous face recognition.
• To address the challenges of small sample size and modality differences in HFR.
• To enhance the performance of HFR systems across diverse imaging scenarios.

Main Methods:

• A novel HFR framework, prototype random subspace (P-RS), is proposed.
• Both probe and gallery images are represented by nonlinear similarities to prototype face images.
• Features are projected into a linear discriminant subspace, and random sampling is incorporated to handle small sample sizes.

Main Results:

• The P-RS method demonstrates improved accuracy in heterogeneous face recognition.
• The framework's effectiveness is validated across four distinct scenarios: NIR to photograph, thermal to photograph, viewed sketch to photograph, and forensic sketch to photograph.
• The integration of nonlinear prototype representation and random subspace projection yields robust performance.

Conclusions:

• The proposed P-RS framework offers a robust and accurate solution for heterogeneous face recognition.
• This approach effectively handles variations in imaging modalities and mitigates the small sample size problem.
• The P-RS method shows significant promise for real-world applications in surveillance and forensics.