Deep Imbalanced Learning for Face Recognition and Attribute Prediction.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Face analysis datasets often suffer from highly-skewed class distributions, with few majority classes and scarce minority instances.
• Traditional deep learning approaches for imbalanced data include class re-sampling and cost-sensitive training.
• The effectiveness of these classic schemes for representation learning on imbalanced data requires systematic validation.

Purpose of the Study:

• To investigate and validate classic strategies for representation learning on class-imbalanced face data.
• To develop a novel deep learning method that learns more discriminative representations from imbalanced datasets.
• To improve accuracy in face recognition and face attribute prediction tasks with skewed class distributions.

Main Methods:

• Conducted extensive and systematic experiments to evaluate existing methods for imbalanced data.
• Proposed a novel approach enforcing inter-cluster margins within and between classes to maintain local data balance.
• Implemented Cluster-based Large Margin Local Embedding (CLMLE) using angular margins on a hypersphere manifold.

Main Results:

• Demonstrated that enforcing inter-cluster margins leads to more discriminative deep representations.
• Showed that CLMLE effectively reduces local class imbalance by creating balanced class boundaries.
• Achieved significant accuracy improvements over existing methods on face recognition and attribute prediction tasks.

Conclusions:

• CLMLE, combined with a k-nearest cluster algorithm, offers a powerful solution for class-imbalanced face analysis.
• The proposed method effectively addresses the challenge of scarce minority class instances in deep learning.
• This approach provides a more balanced and accurate representation learning strategy for real-world face datasets.