Theory Of Feature Extraction And Recognition Of Images And Its Application In Face Recognition

The technology of face recognition is a hot topic in the area of pattern recognition. This dissertation presents the studies of feature extraction of images and the design method of classifier. The main points are as follows:(1) A new formula for solving the statistical uncorrelated discriminant vectors is proposed. A unified solving method is proposed based on the generalization of the statistical uncorrelated discriminant vectors in order to solve the generalized statistical uncorrelated discriminant vectors in both cases of a large number of samples and a small number of samples. The numerical experiments on facial database of ORL show the effectiveness of the proposed method.(2) A series of new methods of feature extraction based on the optimal discriminant analysis are proposed, including the new LDA algorithm based on the spectral decomposition of within-class scatter matrix Sw which is effective when the number of class is small, an improvedalgorithm of optimal set of discriminant vectors based on the SVD which is effective for face recognition, and the kernel Fisher discriminant method (KFDM). Experimental results on ORL show that the KFDM outperforms conventional Fisher discriminant methods in face recognition, however the computational load is much higher than those of conventional algorithms.(3)A series of new methods of feature extraction based on the generalized optimal discriminant analysis are proposed, including the improved algorithm based on Guo's algorithm calculating the generalized optimal set of discriminant vectors, the analytical algorithm of solving generalized optimal set of discriminant vectors which is the complete solution of the generalized optimal set of discriminant vectors, and the new algorithm of feature extraction based on the generalized KL transformation which enjoys the advantaqges of higher recognition rate and the lower computational load. Furthur more, we have proposed the relationship between the generalized KL transformation and the generalized optimal discriminant transformation, a new explanation of generalized KL transformation has been obtained. A concise representation method of between-class scatter matrix and population scatter matrix is proposed, which suits for all the applications of pattern recognition using Fisher criteria.(4) Algebraic feature extraction on the spectro-space has been proposed, which combines the wavelet analysis, wavelet packet analysis and the generalized optimal discriminant analysis.(5) A series of design methods of classifiers are proposed, including the classifier based on the generalized inverse and the probabilistic reasoning method (PRM), a new self-adaptive Kohonen clustering network which overcomes the shortcomings of the conventional clusteringalgorithms, and the fuzzy neural classifier. The experimental study efface recognition is presented based on the combination of multi-feature multi-classifier.(6) This paper proposes a hybrid feature extraction method for face recognition, which is a combination of the eigen matrix, Fisher discriminant analysis, and the generalized optimal set of discriminant vectors. Two standard databases from Yale University and Olivetti research laboratory are selected to evaluate the recognition accuracy of the proposed method. The correct recognition rate of 100% is obtained with the ORL database and Yale database in all the experiments when we choose only one sample for each person. The result is better than those of other literatures. We have also made the experiments of face recognition on the NUST face database. The correct recognition rate of 94% is obtained with the NUST database in the experiments when we choose only one sample for each person. However, if we choose randomly two or more than two NUST faces as the training samples, then the recognition accuracy of 100% is obtained. The results of this method are the best on the ORL database, Yale database, and the NUST database up to now.