Reduced resolution and scale space for dominant feature detection in contours

Analysis of contours or curves at various scales or levels of smoothing (scale space) is an important tool in curve segmentation and feature detection. In this paper we propose the application of the hierarchical discrete correlation algorithm for efficiently calculating and creating a scale space of curves. In conjunction with this procedure we also investigate the use of the reduced resolution or Gaussian pyramid representation of the set of smoothed curves as the basis for initially localizing and detecting features. We also create an approximation to the above algorithms that is computationally less expensive. Finally, we propose a new inter-scale method for curve segmentation and feature detection based on the motion of a curve through scale space. (C) 2000 Pattern Recognition Society. Published by Elsevier Science Ltd. All rights reserved.