In this paper we propose an image corner detector based on the direct curvature scale space (DCSS) technique, referred to as the curvature product DCSS (CP-DCSS) corner detector. After the contours of interested objects are extracted from a real-world image, their curvature functions are respectively convolved with the Gaussian function as its standard deviation gradually increases. By measuring the product of the curvature values computed at several given scales, true corners on the contours can be easily detected since false or insignificant corners have been effectively suppressed. A point is declared as a corner when the absolute value of the curvature product exceeds a given threshold and is a local maximum at the mentioned point. CP-DCSS combines the advantages of two recently proposed corner detectors, namely, the DCSS corner detector and the multi-scale curvature product (MSCP) corner detector. Compared to DCSS, CP-DCSS omits a parsing process of the DCSS map, and hence it has a simpler structure. Compared to MSCP, CP-DCSS works equally well, however, at less computational cost.
Discrete Curvature Representations for Noise Robust Image Corner Detection
