Evidence theory for image segmentation using information from stochastic Watershed and Hessian filtering

In this paper; a new segmentation method is presented. It combines the probability density function of the stochastic Watershed and the Frobenius norm of the Hessian operator under the evidence theory framework. The first step of this method is a classification of the values provided by these two sources of information into five classes. Then, a predefined belief scheme is used to assign masses to pixels in each class. The segmentation result is obtained after beliefs fusion using the Dempster's rule of combination. The method is designed for two-label segmentation, contour and non-contour. Experimental results on a set of images from the Berkeley dataset, shows the ability of this method to yield a good segmentation compared to the given ground truths.