Model-based shape classification using shape-transformation-invariant descriptors

Abstract

ABSTRACT The shape classification methods derived from similarity measures based on the shape-transformation-variant descriptors often require shape normalization/st andardization that involves compli cated computations and contour or code matching schemes. In this paper, we introduce a quantitative similarity measure and a new model-based shape classification method which uses exclusively the shape-transformation-invariant descriptors . This method eliminates all possible variations and potential problems caused by shape transformation, and complicated contour matching and/or shape normalization/standardization procedures. Keywords: Shape similarity, shape classification, shape descriptor, shape transformation 1. INTRODUCTION There are two types of geometric shape descriptors: a. Contour-based descriptors, such as boundary chain codes, length, curvature, bending energy, signature, moment, angular distribution, Fourier descriptors, and other approaches such as Hough transform, mathematical morphology, and neural networks. b. Region-based descriptors, such as area, eccentricity, elongatedness, rect angularity, sphericity, roughness, direction, compactness, convex hull, skeleton, decomposition, and etc. These two types of shape descriptors can be re-classified into the following two categories in relation to shape transformation. The shape transformation, without changing the shape itself, includes scaling, rotation, reflection, inversion, translation, and any combinations of them. a. Shape-transformation-variant descriptors (STVD). This includes chain codes, boundary length, area, signature, moment, and etc. STVDs of a given a geometric shape vary with its transformations. b. Shape-transformation-invariant desc riptors (STID). This includes comp actness, eccentricity, elongatedness, rectangularity, and sphericity. STIDs of a given geometric shape do not vary with its transformations. As shown in many previous studies