Largest Matching Areas for Illumination and Occlusion Robust Face Recognition

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biometrics

Background:

• Face recognition systems face challenges with uneven illumination, partial occlusion, and limited training data.
• Traditional methods often struggle to maintain accuracy under these adverse conditions.

Purpose of the Study:

• To introduce a novel face recognition approach that simultaneously addresses uneven illumination, partial occlusion, and limited training data.
• To enhance the robustness and accuracy of face recognition, particularly in unconstrained environments.

Main Methods:

• A new method employing lighting normalization and occlusion de-emphasis.
• Face recognition based on finding the largest matching area (LMA) at each point, unlike fixed-size local area approaches.
• Novel feature extraction, LMA-based image comparison, and unseen data modeling for robustness.

Main Results:

• Outperformed existing single-training-image methods on extended YaleB and AR face databases for identification.
• Achieved comparable or superior performance to multi-training-image methods.
• Outperformed comparable unsupervised methods on the Labeled Faces in the Wild database for verification.
• Demonstrated competitive performance even with corrupted training images.

Conclusions:

• The proposed LMA-based face recognition method offers significant improvements in handling challenging real-world conditions.
• The approach is robust and effective, even with limited or corrupted training data, outperforming existing state-of-the-art techniques.