Affine curve moment invariants for shape recognition

This paper presents a method of affine curve moment invariants for shape recognition. The proposed method extends affine moment invariants from an area domain to a curve domain. First, a new type of curve moments is defined on a parameterized boundary description of an object. A set of affine moment invariants are then derived based on the theory of algebraic invariants. The proposed moment invariants only deal with pixels on a shape boundary and are mathematically proven to be invariant under arbitrary affine transformations. Furthermore, after a parameterized object boundary description is acquired through curve fitting, the newly defined curve moments can be represented by an integration of the fitting function, so that the computation of moment invariants can be obtained from either one-by-one computation of boundary pixels or a direct integration of the fitting function, which makes the computation efficient while at the same time reducing noise effect. Affine curve moment invariants work well in recognition of affine-deformed objects, which is demonstrated by processing various 2-D shapes under affine transformations. (C) 1997 Pattern Recognition Society.