ECML: An Ensemble Cascade Metric-Learning Mechanism Toward Face Verification

Area of Science:

• Computer Vision
• Machine Learning
• Biometrics

Background:

• Face verification is a challenging fine-grained visual recognition task.
• Improving feature discriminative power is crucial for enhancing face verification performance.
• Metric learning is essential for balancing underfitting and overfitting in feature learning.

Purpose of the Study:

• To propose a novel ensemble cascade metric-learning (ECML) mechanism for face verification.
• To alleviate underfitting using hierarchical metric learning in a cascade manner.
• To resist overfitting by employing ensemble metric learning on feature groups.

Main Methods:

• Developed an ensemble cascade metric-learning (ECML) mechanism.
• Introduced a robust Mahalanobis metric-learning (RMML) method with a closed-form solution.
• Integrated RMML into ECML for a one-pass learning paradigm (EC-RMML).

Main Results:

• The proposed EC-RMML method demonstrates superior performance compared to state-of-the-art metric-learning techniques for face verification.
• The ECML mechanism effectively addresses both underfitting and overfitting challenges.
• RMML avoids computational failures associated with inverse matrix computations in other methods.

Conclusions:

• EC-RMML offers a robust and efficient solution for face verification.
• The ECML framework is adaptable and can be applied to various metric-learning approaches.
• This research advances the field of metric learning for biometric identification.