Discriminative Deep Metric Learning for Face and Kinship Verification

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Metric learning methods for face and kinship verification often rely on single Mahalanobis distance metrics.
• Existing methods struggle to capture the nonlinear manifold characteristics of face image data.

Purpose of the Study:

• To propose a discriminative deep metric learning (DDML) method for face and kinship verification under wild conditions.
• To develop a discriminative deep multi-metric learning approach for enhanced feature robustness and verification accuracy.

Main Methods:

• A DDML method trains a deep neural network for hierarchical nonlinear transformations, projecting face pairs into a shared latent feature space.
• A discriminative deep multi-metric learning method jointly learns multiple neural networks to maximize feature correlation and optimize pair distances.
• The methods focus on reducing intra-class variations and increasing inter-class variations for positive and negative pairs, respectively.

Main Results:

• The proposed DDML and multi-metric learning methods demonstrate acceptable performance in both face and kinship verification tasks.
• Experimental results validate the effectiveness of the learned hierarchical nonlinear transformations and joint multi-metric learning.

Conclusions:

• The developed DDML and multi-metric learning techniques offer a promising advancement for face and kinship verification in challenging, real-world conditions.
• These methods improve feature representation by capturing nonlinear manifold structures and leveraging commonalities across multiple feature descriptors.