Spartans: Single-Sample Periocular-Based Alignment-Robust Recognition Technique Applied to Non-Frontal Scenarios

Area of Science:

• Computer Science
• Biometrics
• Artificial Intelligence

Background:

• Face recognition technology faces challenges with pose variations, occlusions, and limited training data.
• Existing methods often struggle in unconstrained environments, requiring multiple samples or precise alignment.
• The periocular region offers stable and discriminative features less affected by expressions and occlusions.

Purpose of the Study:

• To develop a single-sample, alignment-robust face recognition technique tolerant to pose variations.
• To enhance face recognition accuracy in unconstrained scenarios using periocular information.
• To introduce a novel framework, Spartans, for efficient and effective face matching.

Main Methods:

• Utilized a 3D generic elastic model to generate diverse face images from a single sample.
• Focused on the periocular region, marginalizing out more variable facial areas.
• Employed high-dimensional Walsh local binary patterns for robust feature extraction.
• Learned subject-dependent advanced correlation filters for pose-tolerant subspace modeling.
• Incorporated a coupled max-pooling mechanism to improve performance.

Main Results:

• Achieved 89.69% accuracy on the Labeled Faces in the Wild database, outperforming state-of-the-art methods.
• Demonstrated superior performance in image-restricted and unsupervised protocols.
• Validated effectiveness on Face Recognition Grand Challenge and Multi-PIE databases.
• Showcased advanced correlation filters' superiority in learning pose-tolerant subspaces.

Conclusions:

• The Spartans framework provides a highly accurate and robust solution for single-sample, pose-tolerant face recognition.
• Periocular-based feature extraction combined with advanced correlation filters significantly enhances recognition performance.
• The method offers a promising advancement for real-world, unconstrained face matching applications.