A new approach to curvature estimation and description

A new method for detection and estimation of curvature is presented. The algorithm is implemented in the hierarchical feature pyramid proposed in the GOP concept. Curvature is handled at the second level of the pyramid with an orientation description of the image as input. This complex valued image is convolved and the filter responses are combined into a description of curvature direction, magnitude and curvature/linearity ratio. The procedure resembles in many ways the algorithms for the first level of the pyramid and is a natural extension of these. The method is easy to implement and the tests made show that it performs well and can handle noisy conditions. Some comparisons with other algorithms have been carried out, and the results indicate that the methodology presented here has a number of important advantages over other methods.<
>