Generalized Distance based Matching of Nonbinary Images

We develope a technique to perform e cient and accurate matching to detect the occurence(s) of a template in a scene. The template may describe an object or a textured surface. Representative binary features such as edge points for objects and a set of intensity extrema for textured surfaces are extracted from both test and template images. The search is based on a three-stage \coarsenenal" matching. First, we match with the binarized template skipping positions in horizontal and vertical directions and nd a number of highest ranking matching positions. Second, we search in neighborhoods centered at the stated positions and nd improved matching positions. Finally, some error measure between the original, i.e. gray level or color, test and template images is computed at the ne positions as the nal matching. It is demonstrated that the proposed scheme achieves signi cant reduction in computational complexity with virtually no loss of detection performance.