255 Multi-scale Cortical Keypoint Representation for Attention and Object Detection Jo˜ ao Rodrigues 1 and Hans du Buf 2 1 University of Algarve, Escola Superior Tecnologia, Faro, Portugal 2 University of Algarve, Vision Laboratory, FCT, Faro, Portugal Abstract. Keypoints (junctions) provide important information for focus-of-attention (FoA) and object categorization/recognition. In this paper we analyze the multi-scale keypoint representation, obtained by applying a linear and quasi-continuous scaling to an optimized model of cortical end-stopped cells, in order to study its importance and possibili- ties for developing a visual, cortical architecture. We show that keypoints, especially those which are stable over larger scale intervals, can provide a hierarchically structured saliency map for FoA and object recognition. In addition, the application of non-classical receptive field inhibition to keypoint detection allows to distinguish contour keypoints from texture (surface) keypoints. 1 Introduction Models of cells in the visual cortex, i.e. simple, complex and end-stopped, have been developed, e.g. [4]. In addition, several inhibition models [3, 11], keypoint detection [4, 13, 15] and line/edge detection schemes [3, 13, 14], including dis- parity models [2, 9, 12], have become available. On the basis of these models and processing schemes, it is now possible to create a cortical architecture for figure-background separation [5, 6] and visual attention or focus-of-attention (FoA), bottom-up or top-down [1, 10], and even for object categorization and recognition. In this paper we will focus on keypoints, for which Heitger et al. [4] developed a single-scale basis model of single and double end-stopped cells. W¨ urtz and Lourens [15] presented a multi-scale approach: spatial stabilization is obtained by averaging keypoint positions over a few neighboring micro-scales. We [13] also applied multi-scale stabilization, but focused on integrating line/edge, keypoint and disparity detection, including the classification of keypoint structure (e.g. T, L, K junctions). Although the approaches in [13, 15] were multi-scale, the aim was stabilization at one (fine) scale. Here we will go into a truly multi- scale analysis: we will analyze the multi-scale keypoint representation, from very fine to very coarse scales, in order to study its importance and possibilities for developing a cortical architecture, with an emphasis on FoA. In addition, we will include a new aspect, i.e. the application of non-classical receptive field (NCRF) inhibition [3] to keypoint detection, in order to distinguish between object structure and surface textures. J.S. Marques et al. (Eds.): IbPRIA 2005, LNCS 3523, pp. 255–262, 2005. c  Springer-Verlag Berlin Heidelberg 2005