A bivariate autoregressive technique for analysis and classification of planar shapes

A bivariate autoregressive model is introduced for the analysis and classification of closed planar shapes. The boundary coordinate sequence of a digitized binary image is sampled to produce a polygonal approximation to an object´s shape. This circular sample sequence is then represented by a vector autoregressive difference equation which models the individual Cartesian coordinate sequences as well as coordinate interdependencies. Several classification features which are functions or transformations of the estimated coefficient matrices and the associated residual error covariance matrices are developed. These features are shown to be invariant to object transformations such as translation, rotation, and scaling. Laboratory experiments involving object sets representative of industrial shapes are presented. Superior classification results are demonstrated