Deep Discriminative Feature Models (DDFMs) for Set Based Face Recognition and Distance Metric Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Set-based face recognition faces challenges with intra-class variations.
• Approximating complex face manifolds requires efficient feature representation.

Purpose of the Study:

• To introduce two novel methods for creating compact deep feature models for set-based face recognition.
• To enhance the accuracy and efficiency of face recognition systems using image set approximations.

Main Methods:

• Treating image sets as nonlinear face manifolds composed of linear components.
• Approximating image subsets using deep feature representations (subset centers) learned via distance metric learning.
• Employing discriminative common vectors (projected subset centers) to approximate subsets with an affine hull, removing within-class variances.

Main Results:

• The proposed methods achieve state-of-the-art accuracies on various face recognition and visual object classification datasets.
• The methods effectively approximate image sets using compact deep feature models.
• Distance metric learning with triplet loss on quantized data demonstrates significant advantages over classical methods.

Conclusions:

• The developed methods offer a robust approach to set-based face recognition by effectively modeling face manifolds.
• These techniques provide a powerful alternative to traditional distance metric learning, particularly for quantized data.
• The findings suggest a significant advancement in the field of automated face recognition systems.