Radial Edge Configuration for Semi-local Image Structure Description Lech Szumilas, Horst Wildenauer, Allan Hanbury, and Ren´ e Donner Pattern Recognition and Image Processing Group Vienna University of Technology A-1040, Vienna, Austria lech@prip.tuwien.ac.at http://www.prip.tuwien.ac.at/∼lech Abstract. We present a novel semi-local image descriptor which en- codes multiple edges corresponding to the image structure boundaries around an interest point. The proposed method addresses the problem of poor edge detection through a robust, scale and orientation invari- ant, descriptor distance. In addition, a clustering of descriptors capable of extracting distinctive shapes from a set of descriptors is described. The proposed techniques are applied to the description of bone shapes in medical X-ray images and the experimental results are presented. 1 Introduction Edges are an intuitive way to represent shape information, but the problems asso- ciated with poor edge detection often affect the final result based on edge match- ing or classification. To overcome this problem we introduce a novel semi-local shape descriptor which represents the shape of an image structure by means of edges and their configurations. Our Radial Edge Configuration -descriptor (REC) encodes edges found in a neighborhood of an interest point as a sequence of radial distances in a polar coordinate system (centered on the interest point). Thus, the similarity of shape is assessed by the comparison of local edge configurations. Building on RECs we investigate the possibility to extract groups of similar edge structures by unsupervised clustering of edges. Here, our main contributions are: The definition of a rotation and scale-invariant distance measure between edge configuration descriptors. The most important property of the distance mea- sure is it’s ability to match multiple edges, preserving their spatial relationships and rejecting outlier edge pairs at the same time. This allows for a compari- son of image structures across different scales, with only partially established correspondences. Another particularity of the chosen approach is that scale and orientation are not estimated during descriptor extraction. Instead they are es- tablished as relative entities between two REC descriptors during the distance calculation, which leads to more stable results. The second contribution is the in- troduction of hierarchical edge clustering using this distance. Clustering similar edges in an image is particularly difficult when edges have different lengths or are fragmented due to poor detection. However, the proposed distance measure was G. Bebis et al. (Eds.): ISVC 2007, Part I, LNCS 4841, pp. 633–643, 2007. c  Springer-Verlag Berlin Heidelberg 2007