Correlating multiple redundant scales for corner detection

Corner detection is an important task in computer vision and image processing applications. Basically, corners are high curvature points (HCP), which can be detected by contour analysis. In this paper we propose an approach to detect corners using multiscale analysis. The algorithm provides an undecimated wavelet decomposition of the angulation signal of a shape contour and the high curvature points are identified by correlating multiple redundant scales. The goal is to detect the dominant points of a shape that accurately represent it. Assessment results have shown that the method succeeded in reconstructing the shape contour using the detected HCPs. A novel evaluation measure is also presented in order to confirm that the proposed algorithm outperforms other methods used for testing and comparison purposes. The technique is promising and effective for image retrieval applications.