A Symmetrical Siamese Network Framework With Contrastive Learning for Pose-Robust Face Recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning has advanced face recognition, but performance degrades significantly with pose variations.
• Pose-based long-tailed data, characterized by imbalanced samples of profile vs. near-frontal faces, hinders model training.
• Self-occlusion and texture warping in profile faces complicate discriminative feature learning.

Purpose of the Study:

• To propose a novel framework, Symmetrical Siamese Network (SSN), to address limitations in pose-based long-tailed data.
• To develop a method for learning pose-invariant features in face recognition.
• To improve the robustness of face recognition models against significant pose variations.

Main Methods:

• Introduced the Symmetrical Siamese Network (SSN) with two sub-modules: Feature-Consistence Learning sub-Net (FCLN) and Identity-Consistence Learning sub-Net (ICLN).
• FCLN utilizes contrastive learning to focus on consistent facial areas across simulated pose variations.
• ICLN employs Identity Consistence Loss on profile images to minimize intra-class feature variation across poses.

Main Results:

• The collaborative learning of FCLN and ICLN ensures balanced parameter updates between near-frontal and profile face images.
• Effectively addresses the pose-based long-tailed problem in face recognition datasets.
• Achieved comparable results to state-of-the-art methods on several public face recognition datasets.

Conclusions:

• The proposed SSN framework demonstrates significant improvements in pose-robust face recognition.
• The SSN is adaptable, with LightCNN as a backbone, and compatible with other popular networks.
• This approach offers a viable solution for enhancing face recognition systems in real-world scenarios with diverse poses.