Transforms for multivariate classification and application in tissue image segmentation

Linear and nonlinear transformation techniques were developed for multivariate classification and color image segmentation. It was demonstrated that the Fisher's linear discriminant, which yields a single-dimensional linear transform, results in a loss of class discrimination in certain cases. A generalized multivariate linear transformation technique was thus developed to avoid the undesirable loss of information of class discrimination. Experiments show that this generalized Fisher's linear transformation is effective for classification. Through space augmentation, a nonlinear transformation technique was developed on the basis of the generalized Fisher's linear transformation to extract nonlinear discriminant features for classification and image segmentation. Test results show that this nonlinear transform is capable of extracting latent features to enhance the separability of clusters that are not linearly separable.; An unsupervised image segmentation technique was developed to segment tissue images. A method to determine the initial cluster values broadens the applicability of the segmentation algorithm. Use of nonlinear transforms further enhances the power of the image segmentation algorithms.; The algorithms were implemented by the object-oriented design and programming (OOD/OOP) methodology on the Windows platform for the purpose of beef image segmentation. The meat image processing application was tested with beef images of two different sources captured in different environments. The results demonstrated the effectiveness of the techniques and algorithm developed.