Weighted Fuzzy Approach For Face Recognition

Face recognition, which aims at how to analyze and process face images to extract effective discriminative information by computer and accomplish the identification of individuals in the images, is one of the key technologies of biometric identification. Recently it has been a research focuses in the field of image processing, artificial intelligence, pattern recognition and computer vision et al.During the past decades, researchers attempt to mimic human being's ability of identifying faces, and have presented many effective algorithms, that aimed to improve the average recognition accuracy via different techniques. Nevertheless, the worst recognition accuracy for individual plays an important role for the general performance of the algorithm and usability in practice. This thesis presents a weighted fuzzy linear discriminant analysis approach to improve the worst class recognition accuracy by enlarging the smallest class-between distance in the feature subspace.Fisherface method, which has been successfully applied in the high-dimensional and small sample size problem such as face recognition, is one of the appearance-based algorithms. It extracts face images'feature based on principal component analysis, and has many advantages such as not sensitive to variations in illumination and facial expression. Fisherface method became one of the most popular approaches for dimensionality reduction since the advantages above, and many extensions have been done. Fuzzy linear discriminant analysis refines feature extraction by introducing fuzzy technique, which uses membership to describe the distribution of the samples, to get a better class centers for the train samples. This thesis adopts the idea of fuzzy LDA, and introduces different weighting coefficients on the class-between distances for obtaining the projection matrix. By optimizing the weighting coefficients via genetic algorithm (GA), minimal class-between distance has been enlarged effectively and the projection matrix has been renewed. After projecting the samples into feature space, the linear separability of the training samples can be improved.The proposed algorithm is tested on 3 public face databases, with comparison with Fisherface method and fuzzy linear discriminant analysis.The experiment results reveal that the proposed algorithm enlarges the smallest class-between distance in the reduced subspace effectively, and improve performance of the recognition algorithm significantly.