RVFace: Reliable Vector Guided Softmax Loss for Face Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Deep convolutional neural networks (CNNs) have advanced face recognition.
• Existing margin-based softmax loss functions improve feature discrimination but struggle with noisy labels, lack informative feature mining, neglect non-ground truth class discriminability, and handle class imbalance poorly.

Purpose of the Study:

• To develop a novel loss function for face recognition that addresses the limitations of existing methods.
• To improve feature discrimination by explicitly handling noisy labels and adaptively emphasizing semi-hard features.

Main Methods:

• A unified loss function is proposed that integrates noisy label estimation and dropping.
• The method adaptively emphasizes semi-hard feature vectors from reliable data for discriminative learning.
• This approach combines noisy label detection, feature mining, and feature margin enhancement.

Main Results:

• The proposed loss function effectively addresses issues related to noisy labels and feature learning.
• It achieves more discriminative features for face recognition compared to state-of-the-art methods.
• Extensive experiments on various benchmarks validate the method's effectiveness.

Conclusions:

• The novel loss function offers a significant improvement in face recognition by tackling inherent dataset noise and optimizing feature learning.
• This work represents a pioneering effort in unifying noisy label detection, feature mining, and feature margin strategies within a single loss function for enhanced face recognition performance.