Behavior Analysis And Detection Research Of Corner In Contour Difference Of Gaussian Scale Space

Corner is one of the most important feature in images and has been widely used invarious fields of computer vision and image processing. This has made corner detectiona fundamental problem in computer vision and image processing. The introduction ofscale space theory and multi-scale technique has played an important role in thedevelopment of corner detection method. This paper will investigate the behavior ofcorners in Difference of Gaussian Scale Space with the same method as in the study ofbehavior of corners in Curvature Scale Space.DoG map and the unified corner model are firstly defined to simplify the study.DoG scale space of the unified corner model is constructed by convolving the DoGoperator with the unified model at different scales and the implicit expression of DoGmap of the unified corner model is derived. Then several properties of the behavior ofcorners in the scale space are obtained. When the scale increases, the trajectory ofcorner in DoG scale space will persist, vanish or merge. The trajectories of adjacentcorners attract or repel with each other. Finally, based on a numerical method of theDoG operator, the behavior of corners in the discrete DoG scale space is investigated.On the basis of the behavior of corners in the DoG scale space, the idea ofmulti-scale product is introduced to improve the DoG corner detection algorithm and amulti-scale product DoG corner detection algorithm is proposed. In order to assess theimprovement of the new algorithm, several experiments are made. Invariance test androbustness test verified the effectiveness of the algorithm and the comparativeexperiments based on the correct detection rate and localization error demonstrate theperformance improvement.