Subject-specific and pose-oriented facial features for face recognition across poses.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biometrics

Background:

• Traditional face recognition often relies on frontal or mug shot images for enrollment.
• Recognizing faces from varied poses, especially when enrollment images differ significantly, remains a challenge.
• Forensic applications typically assume mug shots are available, but surveillance scenarios capture diverse poses.

Purpose of the Study:

• To develop a robust face recognition method for scenarios with disjoint sets of facial poses for enrollment and recognition.
• To address the challenge of recognizing individuals when enrollment and probe images exhibit significant pose variations.
• To create a system capable of identifying faces across different viewpoints without requiring frontal images during enrollment.

Main Methods:

• Feature extraction involves clustering enrollment poses and decomposing facial appearance using Embedded Hidden Markov Models (EHMM).
• Subject-specific and pose-oriented (SSPO) facial components are defined for each individual.
• Classification utilizes an Adaboost weighting scheme to fuse SSPO component classifiers.

Main Results:

• The proposed method demonstrates superior performance compared to existing approaches.
• Outperformed a component-based classifier using manually cropped local facial features.
• Extensive performance evaluation confirmed the effectiveness of the SSPO component approach.

Conclusions:

• The developed method effectively handles face recognition with disjoint pose sets, offering a significant advancement.
• The SSPO components derived from EHMM provide discriminative features for pose-invariant face recognition.
• This approach offers a promising solution for real-world surveillance and forensic applications with unconstrained face poses.