Related Experiment Videos
GA-fisher: A new LDA-based face recognition algorithm with selection of principal components
Wei-Shi Zheng1, Jian-Huang Lai, Pong C Yuen
1Mathematics Department, Sun Yat-sen University, Guangzhou, China. SunnyWeiShi@163.com
Summary
This study introduces a novel Genetic Algorithm-based Fisherface (GA-Fisher) method for face recognition. GA-Fisher systematically selects optimal eigenvectors for dimension reduction, improving accuracy and efficiency over traditional methods.
Area of Science:
- Computer Vision
- Machine Learning
- Pattern Recognition
Background:
- The Fisherface method for face recognition faces challenges with singular within-class scatter matrices (Sw) when training samples are fewer than image dimensions.
- Principal Component Analysis (PCA) is often used for dimension reduction, but selecting optimal eigenvectors for Linear Discriminant Analysis (LDA) remains unsystematic.
- Existing approaches for handling illumination effects by removing eigenvectors lack a systematic basis.
Purpose of the Study:
- To develop a systematic and automatic method for selecting eigenvectors in Fisherface for improved face recognition.
- To address the dimension reduction problem in Fisherface, particularly when the within-class scatter matrix (Sw) is singular.
- To enhance the computational efficiency and accuracy of face recognition systems.
Main Methods:
- A theorem is proposed to interpret the use of PCA within LDA.
- A Genetic Algorithm (GA) is employed to develop a systematic method for selecting eigenvectors, resulting in GA-PCA (Genetic Algorithm-Principal Component Analysis).
- A GA-Fisher method is developed, incorporating GA-PCA for dimension reduction and a whitening procedure to improve computational efficiency.
Main Results:
- The GA-PCA approach demonstrates that smaller eigenvectors are crucial for effective dimension reduction.
- The GA-Fisher method achieves approximately 5% performance improvement compared to the traditional Fisherface method.
- Experimental evaluations on FERET and CMU PIE databases validate the effectiveness of the proposed method.
Conclusions:
- The GA-Fisher method provides an optimal basis for dimensionality reduction through GA-PCA.
- The systematic selection of eigenvectors and the whitening procedure enhance both accuracy and computational efficiency.
- The proposed method offers a significant advancement in Fisherface-based face recognition systems.