Gabor wavelet similarity maps for optimising hierarchical road sign classifiers Alan Koncar a, * , Holger Janßen b , Saman Halgamuge a a Dynamic Systems and Control Group, Department of Mechanical and Manufacturing Engineering, University of Melbourne, VIC 3010, Australia b Research and Development, Robert Bosch GmbH, Hildesheim, Germany Received 5 January 2006; received in revised form 17 July 2006 Available online 8 September 2006 Communicated by M. Kamel Abstract In recent years it has been shown that hierarchical classifiers have a significant advantage over single stage classifiers both in classi- fication accuracy and in complexity of the classification features. This paper introduces a new method for creating the structure of hier- archical classifiers using a novel method for determining clusters. The proposed method uses features obtained using Gabor wavelets to create similarity maps, which help separating the class space into smaller more distinctive clusters. This approach has been applied on the Road Sign Recognition problem and has shown encouraging results in comparison to k-means algorithm. Ó 2006 Elsevier B.V. All rights reserved. Keywords: Gabor wavelets; Jets; Euclidean distance; Normalised scalar product; Hierarchical classifier; Gabor similarity maps; Road sign recognition 1. Introduction The method of using many classifiers combined in a hierarchical structure in order to improve overall classifica- tion performance has been studied extensively in recent years (Priese, 1994; Paclik and Novovicova, 2000; Rodri- guez et al., 2002; Mathis and Bruel, 2002; Sung et al., 2005; Joshi et al., 2006; Giusti et al., 2002; Mayraz and Hinton, 2002; Heisele et al., 2003; Schikuta, 1996; Gdal- yahu and Weinshall, 1999; Zin and Hama, 2004; Shap- oshnikov, 2002; Zadeh, 1997). This method is based on the idea to delay the final classification decision by intro- ducing several stages into the classification process. At the highest level of the hierarchy a set of features is used to distinguish groups of classes (clusters) from each other. In lower levels of hierarchy the classification decision relies on a different set of features that is more appropriate for a given group of classes. Features used in different classification stages may vary from simple grey values, colour information and/or geo- metrical shape (Priese, 1994; Mayraz and Hinton, 2002; Gdalyahu and Weinshall, 1999; Shaposhnikov, 2002; Zadeh, 1997; Escalera et al., 1997), image histograms (Pac- lik and Novovicova, 2000), Haar features with AdaBoost (Heisele et al., 2003), features derived using Gabor filters (Ka ¨ma ¨ra ¨inen et al., 2002; Daugman, 1985; Wiskott et al., 1997; Wang et al., 2002; Hamamoto et al., 1996; Sung et al., 2005; Joshi et al., 2006) and many others. Furthermore the hierarchical approach to classification has been applied to many problems in pattern recognition such as road sign recognition (Priese, 1994; Paclik and Novovicova, 2000; Zin and Hama, 2004; Shaposhnikov, 2002; Zadeh, 1997), face detection and recognition (Wis- kott et al., 1997; Heisele et al., 2003), handwritten numerals (Hamamoto et al., 1996; Sung et al., 2005; Mayraz and Hinton, 2002), characters of various language scripts 0167-8655/$ - see front matter Ó 2006 Elsevier B.V. All rights reserved. doi:10.1016/j.patrec.2006.07.012 * Corresponding author. Tel.: +61 3 8344 6877; fax: +61 3 9347 8784. E-mail addresses: koncar.alan@gmail.com (A. Koncar), holger.jans- sen@de.bosch.com (H. Janßen), saman@unimelb.edu.au (S. Halgamuge). www.elsevier.com/locate/patrec Pattern Recognition Letters 28 (2007) 260–267