SFace: Sigmoid-Constrained Hypersphere Loss for Robust Face Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep face recognition models achieve high performance using large datasets and advanced loss functions.
• Current methods focus on minimizing intra-class and maximizing inter-class distances, potentially over-optimizing noisy data.
• Imperfections in training datasets can lead to overfitting in deep face recognition models.

Purpose of the Study:

• To propose a novel loss function, Sigmoid-Constrained Hypersphere Loss (SFace), to address overfitting in deep face recognition.
• To optimize intra-class and inter-class objectives moderately, considering the imperfections of training databases.
• To enhance the robustness of deep face recognition models against noisy training data.

Main Methods:

• Introduction of the Sigmoid-Constrained Hypersphere Loss (SFace) function.
• Implementation of intra-class and inter-class constraints on a hypersphere manifold.
• Utilizing sigmoid gradient re-scale functions to control optimization degrees for training samples.

Main Results:

• SFace effectively balances the reduction of intra-class distances for clean samples with the prevention of overfitting to label noise.
• Models trained with SFace demonstrate improved robustness in deep face recognition.
• Experiments on large-scale datasets (CASIA-WebFace, VGGFace2, MS-Celeb-1M) and benchmarks (LFW, MegaFace, IJB-C) confirm SFace's superiority.

Conclusions:

• The proposed SFace loss function offers a more robust approach to deep face recognition by handling imperfect training data.
• SFace contributes to developing more reliable and accurate face recognition systems.
• The method shows significant improvements across various standard face recognition benchmarks.