Localization of dominant points for object recognition: A scale-space approach

The dominant points which are used for identifying an unknown object shape are commonly taken as the positions where the maxima of the curvature function occur on the object boundary. The peaks in the curvature function can be deduced from the corresponding zero-crossing points of its first-order derivative. In practice, presence of noise signal in the object contour may introduce false zero-crossings in the differentiation process, resulting in the apparent existence of false dominant points. Attenuation of the noise signal can be realized by convolving the contour function with a Gaussian filter. The width of the Gaussian function, however, has to be properly decided to prevent unnecessary removal of the relevant dominant points. In this paper, a novel scheme for automatic determination of the filtering scale is reported. The method employs scale-space decomposition to form a basis for an explicit and quantitative measurement on the reliability of the dominant point sets detected under different degree of filtering, with which the one exhibiting the highest score is selected. The method has been successfully applied to extract the dominant point sets for different types of handtools without prior knowledge of their sizes, shapes and orientations.