Multiscale image segmentation based on one class SVM and wavelet

Abstract

A supervised multiscale image segmentation method is presented based on one class support vector machine (OCSVM) and wavelet transformation. Wavelet coefficients of training images in the same directions at different scale are organized into tree-type data as training samples for OCSVMs. Likelihood probabilities for observations of segmentation image can be obtained from trained OCSVMs. Maximum likelihood classification is used for image raw segmentation. Bayesian rule is then used for pixel level segmentation by fusing raw segmentation result. In experiments, synthetic mosaic image, aerial image and SAR image were selected to evaluate the performance of the method, and the segmentation results were compared with presented hidden Markov tree segmentation method based on EM algorithm. For synthetic mosaic texture images, miss-classed probability was given as the evaluation to segmentation result. The experiment showed the method has better segmentation performance and more flexibility in real application compared with wavelet hidden Markov tree segmentation.