Self-training-based face recognition using semi-supervised linear discriminant analysis and affinity propagation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Face recognition is a critical application in computer vision and machine learning.
• Traditional supervised methods demand extensive labeled face image datasets, which are often scarce in real-world scenarios.
• Unlabeled face image data is frequently abundant, presenting an opportunity for alternative learning paradigms.

Purpose of the Study:

• To develop an effective semi-supervised face recognition method that addresses the scarcity of labeled data.
• To integrate semi-supervised linear discriminant analysis (SDA) and affinity propagation (AP) within a self-training framework.
• To enhance face recognition performance by leveraging both labeled and abundant unlabeled face images.

Main Methods:

• A self-training framework was developed, integrating semi-supervised linear discriminant analysis (SDA) and affinity propagation (AP).
• SDA was utilized to compute a discriminative face subspace incorporating both labeled and unlabeled images.
• AP identified exemplars for different face classes within the computed subspace, facilitating iterative data labeling and model refinement.

Main Results:

• The proposed semi-supervised face recognition algorithm demonstrated superior performance across four diverse face datasets.
• Experimental results indicated that the integrated SDA and AP approach outperformed existing unsupervised, semi-supervised, and supervised methods.
• The self-training framework effectively utilized unlabeled data to improve recognition accuracy.

Conclusions:

• The developed semi-supervised face recognition method offers a robust solution for scenarios with limited labeled data.
• The integration of SDA and AP within a self-training framework is a promising direction for advancing face recognition technology.
• This approach significantly enhances face recognition performance by effectively exploiting readily available unlabeled data.