Template-Driven Knowledge Distillation for Compact and Accurate Periocular Biometrics Deep-Learning Models.

Area of Science:

• Computer Science
• Biometrics
• Machine Learning

Background:

• Building accurate periocular recognition models with limited parameters is challenging.
• Deeper models excel at learning complex information but are computationally expensive.
• Knowledge distillation (KD) transfers knowledge from large teacher models to smaller student models.

Purpose of the Study:

• To develop a novel template-driven KD approach for periocular recognition.
• To enhance the generalizability and accuracy of small periocular recognition models.
• To improve biometric template generation for verification and storage.

Main Methods:

• Proposed a template-driven KD method optimizing student model template generation to match teacher model templates.
• Evaluated the approach on intra- and cross-device periocular verification tasks.
• Compared performance against models trained without KD and with conventional KD.

Main Results:

• The template-driven KD approach significantly outperformed conventional KD and no-KD methods.
• Achieved a notable reduction in Equal Error Rate (EER) for cross-device verification (14.7% vs. 21.9% and 22.2%).
• Demonstrated the effectiveness of distilling template-level knowledge for improved biometric performance.

Conclusions:

• Template-driven KD is a superior method for training accurate and generalizable small periocular recognition models.
• This technique offers a more efficient way to achieve high performance in biometric systems.
• The proposed method advances the field of biometric template protection and recognition.