Based On The Matrix Model And The Quantum Mode Feature Extraction And Its Classification

Feature extraction is a critical problem in the research fields of pattern recognition. It can affect the design and performance of classifiers extensively. The basic mission of feature extraction is to identify a set of features that are most effective for subsequent classification task from a set of candidate features. The state of the art in feature extraction methods includes statistics-based methods (e.g. PCA & FLDA), knowledge-based methods and neural-networks based methods, etc. In this thesis two novel classes of feature extraction methods are proposed, i.e. matrix-pattern-based and vector subpattern-based representation methods respectively.Pervious research in PCA and FLDA mainly focuses on vector-based pattern representation methods for feature extraction and dimensionality reduction, i.e. all patterns must be transformed into vector before any subsequent processing. So patterns represented in matrix form (e.g. an image) must be stretched into vectors in preprocessing step. This type of strategy has two main shortcomings: 1) useful information for classification task contained in the matrix structure may be jeopardized in the vectorizing procedure; 2) after vectorizing procedure computation complexity in classification task may increase substantially due to the vector pattern representation. Our proposed methods, MatPCA & MatFLDA, can deal with not only the vector pattern, but also matrix pattern. In addition, a vector pattern can be recombined into a matrix pattern using some matrixization technique and then be processed by MatPCA & MatFLDA.In above mentioned MatPCA & MatFLDA, a vector pattern is firstly reshaped into a matrix pattern and then processed by PCA & FLDA. It follows a first-matrixization-then-extraction path. What will happen if the "matrixization" is substituted by "partition"? That is of our interest and is explored in the thesis's second part. We proposed SpPCA & SpFLDA feature extraction methods. Contrary to MatPCA & MatFLDA, they follow a first-partition-then-extraction procedure. In these two methods a vector pattern is firstly partitioned into a set of sub-patterns, i.e. each sub-pattern in this set is only a part of the original vector pattern. After the partition, traditional PCA & FLDA are used on these sub-pattern sets for sub-feature extraction. Finally the obtained sub-feature combination of an original pattern is used to subsequent classification.Experimental results relating to the efficiency and accuracy of these proposed methods are also presented in this thesis.