A new method for the automatic segmentation of color image is proposed in this thesis. As a starting point, it is assumed that the observed distribution of the color feature vectors is a mixture of several multivariate Gaussian distributions with unknown mean vectors, covariance matrices, and a priori probabilities. For the purpose of determining the number of classes and their Gaussian parameters, the partitioning of the color feature space into pararellepiped spaces is acomplished by thresholding the intensity histogram of each color component. It is determined by significance test that each pararellepiped space includes the significant class. The classification of color feature vectors is carried out according to the modified minimum-error-rate rule. This classification is iterated until the states of all the pixels become the classified states. Each pixel is labelled as the classified state when the class label by the previous classification is the same as the class label by the current classification. For each iteration, the Gaussian parameters are updated for each class. Using the proposed method, we can obtain the good segmentations for the several color images.
