Independent Component Analysis In The Feature Extraction Of Pattern Classification

This thesis is a part of the research project (0339037)"Establish Assistant Diagnostic Pattern Classifiers Based On Hair Mineral Elements Analysis"of the Education Department and Scientific Foundation of Guangxi Zhuang Automous Region. The aim of this research project is to establish assistant diagnostic pattern classifiers by revealing the relationship between hair mineral element concentration levels and diseases using statistical pattern recognition methods. It is common that there exist some correlations among the different features of the samples, so, one shall use some appropriate feature extraction method to get the most suitable features. In the progress of establishing a pattern recognition system, feature selection and feature extraction are the core jobs due to they are closely related to the design of classifying algorithms and are the main factors affected the performance of designed classifiers. Classical feature extraction methods include: Principle Component Analysis,Singular Value Decomposition,Projection Pursuit,Self-Organizing Map, and so on. However, the inherited drawback of those feature extraction methods is limited to second-order statistics only. The major work of this thesis lies in using a new feature extraction method----Independent Component Analysis. The principle of independent component analysis algorithm is to find the mutual independent underlying components, to remove the higher-order redundant between components and to extract the independent original signals according to the analysis of higher-order statistical relationships among the multidimensional observed data. Features extraction by ICA mainly aims at natural images processing. And in the thesis, ICA is planned to extract potential features in the data of hair mineral element concentration levels of Naso-pharyngeal Cancer patients. The specific work as flowering: (1) Process the data of hair mineral element concentration levels of Naso-pharyngeal Cancer patients, and analysis the data. (2) Analysis and explain in detail the independent component analysis theory and algorithm. (3) Use FastICA algorithm,Infomax algorithm based on information theory respectively to carry out feature extraction of the processed data. (4) Use Linear Discrimination Classifier,Quadratic Bayes Normal Classifier,Optimization of the Parzen classifier,K-Nearest Neighbor Classifier to classify the data. (5) Analysis : the experimental result shows that ICA can extract effectly the best features from the multidimensional data of nonlinear classification problem, and deserves more and thorough investigations.