THE ASSIGNMENT OF TOPOGRAPHIC LABELS BY A STATISTICAL MODEL-BASED APPROACH

A rich and robust way to understand an image in computer vision system is the topographic primal sketch proposed by Haralick et al. (1983). The two aspects that originate there and need further attention are (1 ) the choice of the method for surface function estimation, and (2) us e of topographic labels for the purpose of segmentation of tonal image s. Investigation of the first aspect is reported in this work. We prop ose a function basis which is shown to be optimum and devise a method for the surface function estimation from a noisy image using this basi s. We then implement the topographic classification scheme of Haralick et al. (1983), and in doing so, we use the noise variance estimate ob tained in surface fitting to decide the significane and sign of variou s quantities like the directional derivatives and associated eigenvalu es, thus removing the arbitrariness in their method of choosing thresh olds. The results of our methodology are illustrated using synthetic a nd natural images and are compared with Haralick's fixed cubic polynom ial method.