Örüntü tanımada yeni altuzay yaklaşımları

In this thesis, three topics of pattern recognition namely, feature selection, single image per subject problem, and within-class scatter matrix estimation of classes are discussed. One of the important factors that affect the performance of the pattern recognition system is the dimension of the feature vector. A novel feature selection method is proposed which is related to DCVA in order to overcome the high dimensionality problem encountered in recognition issues. The important features are determined by the column norms of the projection matrix to the range space of all common vectors. Another factor that affects the performance of the pattern recognition system is the training sample size. Traditional methods which use the within-class scatter matrix fail if one sample from each subject is available because the within-class scatter matrices are all zero. An image decomposition method that uses QR-decomposition with column pivoting (QRCP) is proposed to overcome one sample per class problem. Also a two-dimensional extension of DCVA is proposed. The third and also important problem is to make a good estimation of within-class scatter matrices. But in high dimensional classification problems generally it is not possible to find a sufficient number of samples per class. In our proposal, at first the data is projected onto the range space of the total within-class scatter matrix. Then the within-class scatter matrix of a class is modeled using not only its own data but also the data of all other classes.