Multi-scale curvature based image corner detection and matching

Abstract

Corners are important local features of image. Corner detection and matching techniques play important roles in image understanding and computer vision. The paper proposes a new adaptive corner detection method based on multi-scale curvature representation. First, the contour of the image is extracted by the Canny edge detection technique. The curvature of each point in the contour is calculated at a high scale in Gaussian multi-scale space and the points with local maximum of values of curvature are obtained. An adaptive block technology is proposed to determine candidate corners. Then, the accurate locations of the corners are determined at a low scale level. In order to match corners in images of distinct views, the gradient histogram of neighborhood region around each corner is constructed to determine a principal orientation. A corner matching algorithm based on principal orientation and gray correlation is proposed. Experimental results show that both of new algorithms are stable, reliable, and efficient.