A normalized generalized curvature scale space for 2D contour representation

Abstract

Here, we intend to propose a discrete normalization of the Generalized Curvature Scale Space (GCSS). The GCSS is an Euclidean invariant planar contour descriptor. It consists on the convolution of the contour by Gaussian functions with different scales. The points having the same curvature values as the selected extremums are the considered key points. This representation implies different number of descriptors from a shape to another. Thus, a step of redistribution of the key points is requested. Therefore, a discrete normalization approach is proceeded. The descriptor is composed by the curvature variation of the key points at the smoothed curve. Several datasets were used to carry on the experiments and to verify the accuracy, the stability and the robustness of the novel description. The Dynamic Time Warping distance is the similarity metric used. Experimental results show that considerable rates of image retrieval are reached comparing to the state of the art.