SHAPE RECOGNITION AND DESCRIPTION: A COMPARATIVE STUDY

Abstract : Several methods of analyzing binary images using global shape methods based upon functional approximation have been reported in the literature. However, there has been a lack of information comparing the effectiveness of these methods in shape analysis. Five methods of global shape analysis are compared on two bases: (1) Fourier descriptors of the boundary; (2) Walsh points of the boundary; (3) the cumulative angular deviant Fourier descriptors; (4) moments of the silhouette; and (5) moments of the boundary. First, the different methods are introduced, their geometric properties presented, and the formulae for some generic shapes are provided. The methods are compared on the basis of the empirical facts derived from a set of aircraft shape recognition experiments. The shapes are different views of six aircraft. The aircraft silhouettes are the two-dimensional projected images of three-dimensional rigid bodies. The five methods are ranked according to their performance from these experiments. A new method for the recognition of partial shapes based on the Fourier-Mellin transform is introduced. A shift and scale invariant correlation of the complete and partial shape's curvature functions is obtained by applying the Mellin transform to the magnitude of the Fourier transform of the curvature functions. The logarithm of the shift in the correlation function corresponds to the time scaling of the partial shape's curvature function. Then the shift in an ordinary correlation of the complete shape's curvature is the time shift necessary to complete the alignment. Then a pointwise comparison of the curvature functions can be made to determine matching and non-matching contour segments.