Enlarge the training set based on inter-class relationship for face recognition from one image per person

Area of Science:

• Computer Science
• Artificial Intelligence
• Biometrics

Background:

• Large-scale face recognition tasks, like driver's license identification and law enforcement, often face the "one sample problem," where training datasets contain only a single image per individual.
• Existing face recognition methods typically require multiple images per person for training, rendering them ineffective for the one sample problem.

Purpose of the Study:

• To investigate the limitations of Principal Component Analysis (PCA), Fisher Linear Discriminant Analysis (LDA), and Locality Preserving Projections (LPP) in handling the one sample problem.
• To identify and explain the inherent difficulties of the one sample problem, including small sample size, lack of representative samples, underestimated intra-class variation, and overestimated inter-class variation.

Main Methods:

• Analysis of why standard face recognition techniques like PCA, LDA, and LPP fail with single samples per person.
• Proposal of a novel method to enlarge the training set by utilizing inter-class relationships.
• Extension of LDA and LPP algorithms to extract features from the augmented training data.

Main Results:

• Demonstration of the ineffectiveness of PCA, LDA, and LPP in one sample scenarios.
• Validation of the proposed method's effectiveness in improving face recognition accuracy through experimental results.

Conclusions:

• The one sample problem presents significant challenges due to data limitations and variations.
• Enlarging the training set based on inter-class relationships, coupled with extended LDA and LPP, offers a viable solution for accurate face recognition in low-sample scenarios.