An improved threshold selection method for image segmentation

An unsupervised optimal multi-threshold selection scheme for image segmentation is presented. This method is clearly an improvement on two existing methods, namely, Otsu's (1979) optimal multi-threshold method and Wang's (1991) threshold hierarchy method. The histogram is divided into different classes using interval tree structure by thresholding the histogram at different scale levels /spl sigma/ of Gaussian convolution. The different histogram dominant modes are then fitted by a Gaussian distribution and the intersection of these Gaussian curves are new threshold points for the image.<
>