Enhancing Face Recognition With Detachable Self-Supervised Bypass Networks.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep learning and large datasets have advanced automatic face recognition (FR) to human-level performance.
• Current FR systems struggle with unconstrained scenarios, particularly large pose variations and occlusions.

Purpose of the Study:

• To propose a novel Bypass Enhanced Representation Learning (BERL) method for improving FR robustness in unconstrained settings.
• To enhance feature learning for face recognition by integrating self-supervised and supervised learning techniques.

Main Methods:

• Developed a BERL method incorporating two auxiliary bypasses: 3D reconstruction and blind inpainting.
• The 3D reconstruction bypass aids in learning pose-independent facial features.
• The blind inpainting bypass assists in capturing facial context for occlusion robustness.

Main Results:

• The BERL method was trained end-to-end using self-supervised tasks and supervised face identification.
• Experimental results on multiple benchmarks demonstrate superior performance compared to state-of-the-art methods.
• The learned representations showed strong generalization to downstream tasks like facial attribute recognition.

Conclusions:

• The BERL method effectively enhances face recognition accuracy under challenging conditions (pose, occlusion) without additional annotations or computational overhead.
• The proposed approach offers a robust and efficient solution for unconstrained face recognition and related tasks.