Generalized distance based matching of nonbinary images

We develop a technique to perform efficient 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 "coarse-fine-final" matching. First, we match with the binarized template skipping positions in horizontal and vertical directions and find a number of highest ranking matching positions. Second, we search in neighborhoods centered at the stated positions and find improved matching positions. Finally, some error measure between the original, i.e. gray level or color, test and template images is computed at the fine positions as the final matching. It is demonstrated that the proposed scheme achieves significant reduction in computational complexity with virtually no loss of detection performance.