Texture segmentation using semi-supervised support vector machine

Abstract

Support vector machine (SVM) is used here to detect the texture boundaries. In order to do that, a cost function is initially defined based on the estimation of higher order statistics (HOS) of the intensities within small regions. K-mean algorithm is used to find the centres of the two clusters (boundary or texture) from the values of the cost function over the entire image. Then the target values are assigned to the class members based on their Euclidean distances from the centres. A supervised nonlinear SVM algorithm with RBF kernel is later used to classify the cost function values. The boundary is then identified in places where the cost function has greater values. The overall system will be semi-supervised since the targets are not predetermined; however, the number of classes is considered as two. The results show that the algorithm performance is superior to other conventional classification systems for texture segmentation. The displacement of the edges is negligible.