Discriminative multimanifold analysis for face recognition from a single training sample per person

Area of Science:

• Computer Vision
• Machine Learning
• Biometrics

Background:

• Conventional face recognition relies on multiple samples per person (MSPP), which is often unavailable in real-world scenarios like law enforcement and e-passports.
• The single sample per person (SSPP) challenge limits the performance of many existing face recognition algorithms due to insufficient data for discriminant learning.

Purpose of the Study:

• To develop a novel method for robust face recognition under the single sample per person (SSPP) constraint.
• To address the limitations of traditional methods by learning discriminative features from limited data.

Main Methods:

• Proposed Discriminative MultiManifold Analysis (DMMA) by partitioning face images into non-overlapping patches.
• Formulated SSPP face recognition as a manifold-to-manifold matching problem.
• Learned multiple DMMA feature spaces to maximize inter-person manifold margins and utilized a reconstruction-based manifold-to-manifold distance for identification.

Main Results:

• Demonstrated the effectiveness of the DMMA method on three standard face databases.
• Achieved improved face recognition performance in SSPP scenarios compared to existing approaches.

Conclusions:

• The proposed DMMA method offers a viable solution for face recognition systems with limited enrolled samples.
• The approach effectively handles the challenges posed by the single sample per person (SSPP) problem in practical biometric applications.