Curve segmentation by continuous smoothing at multiple scales

Abstract

In the case of curve description, the accurate and precise estimation of corners plays an important role. Corners exhibit a high change of curvature /spl kappa/', so we propose the location of them by means of the estimation of /spl kappa/' at different scales of smoothing. The main problem arises when the curvature is computed because of the inherent quantization. We propose to calculate the curvature of contours detected at subpixel level so that, the quantization error is reduced. Once the change of curvature has been estimated, the curve is segmented at high /spl kappa/' points. It is important to emphasise the fact that the curves are segmented before a smoothing at coarser scale is applied, otherwise two close corners could be merged as smoothing increases.