Profile Scale-Space for Image Segmentation

Object boundaries in images often exhibit a complex greylevel appearance, and segmentation which accurately fits the target image requires a robust, multiscale statistical model of image appearance around the object. Objects whose appearance differs from region to region call for an image appearance model which is tied to the geometric representation of the object, after the fashion of Active Appearance Models [1]. This paper aims at improving the image match component of a model-based segmentation framework. We extract 1D profiles normal to the boundary, a la Active Shape Models, and develop a scale-space on the profiles, where blurring is done only parallel to the boundary. A statistical model is built on features in the profile scale-space, incorporating weighted feature selection. This yields image forces which, when coupled with shape constraints, provide a Bayesian image segmentation framework.