A Less-Dependent Threshold Corner Detection Algorithm

Existing corner detection algorithms excessively depend on the selection of thresholds, which result in complex implementation and unstable performance. Our purpose is to weaken the influence of threshold in corner detection. This paper presents a less-dependent threshold corner detection method to extract geometrically important corners for image vectorization. In general, the performance of the existing corner detection algorithm depends largely on one or more thresholds. Specifically, corner candidates are extracted with an improved intensity-based detector and contour-based detector, respectively. The corner candidates are classified into two groups: uncertain corners and certain corners which are sorted by a matching method based on Euclidean distance. With the certain corners, we define Self-confident Level (SCL), which does not depend on any threshold, to remove false corners from the uncertain corners. Experiments validate the good performance of the proposed method.