Visual Feature Array based Cognitive polygon recognition using the UFEX text categorizer Barna Resk´ o 1,2,3 , Domonkos Tikk 2 , Hideki Hashimoto 1,4 and P´ eter Baranyi 3,4 1 Institute of Industrial Science, University of Tokyo, Japan 2 Budapest University of Technology and Economics, Dept. Telecommunication and Media Informatics 3 Computer and Automation Research Institute, Hungarian Academy of Sciences, Hungary 4 Integrated Intelligent Systems Japanese–Hungarian Laboratory Abstract— This paper presents a cognitive vision based ap- proach to recognize polygons on a natural image. The approach is based on the Visual Feature Array (VFA), which is a cognitive computational model of the mammalian primary visual pro- cessing. VFA, as a multidimensional orthogonal data structure, contains data about the line segment and vertex features in the edge detected input image. Based on the features available in VFA, using the Universal Feature Extractor classifier (UFEX), the problem of polygon categorization and recognition is addressed. The results are compared to solutions by conventional neural networks, such as the Learning Vector Quantization network. I. I NTRODUCTION Neurobiology and cognitive psychology inspired methods and algorithms and models with the ultimate goal of object recognition are getting more and more researched today [1], [2]. In this paper we propose a method for polygon recognition based on models of cognitive informatics. Polygon recognition may be a step to support more complex object recognition, which in today’s computer vision research is a hot topic. There are several research directions, such as the use of statistical modeling [3], [4] approaches. Neurobiology and cognitive psychology inspired methods, algorithms and models with the ultimate goal of object recognition are also getting more and more researched today [1], [2]. It is claimed that the brain processes the input, represents the knowledge and generates the responses through several hierarchies of abstraction [5]. In this context, the recognition of polygons starts by the recognition of its components (lines and corners). Similarly, the polygons can be used to recognize objects that are composed of them. Other polygon recognition techniques use a geometrical approach. A classical solution for shape recognition is the Hough transform [6], which requires a parametric description of the shape that is sought for. The recognition is made by scanning through the parameter space describing the object, and looking for matches on the input image. The Hough transform performs well on simple shapes (lines or circles), but computational complexity increases exponentially with the number of parameters describing the shapes. The hough transform also examines every pixel of an object, thus requires a large amount of computational capacity. The polygon recognition proposed in this paper builds on the information stored in the visual feature array (VFA), proposed in [7]. VFA is a multidimensional array of indicators, as an analogy of the neurons of different functionality in the visual cortex. An indicator (from here we refer to it as a VFA element or element) is positioned in VFA according to its parameters, which describe its functionality. A high value of an element means that a feature with the associated parameters is present in the input image. Such a representation of the image information allows to find and recognize more complex object on the image, which is not composed of pixels, but features, such as lines and vertices. The present method takes the information available in VFA, and groups them to become the components of a hypothetical polygon. The grouping is performed heuristically, e.g. based on the proximity of similar or different features. The hypothesis is verified using UFEX (Universal Feature Extractor), a high performance categorizer, used mainly for text categorization problems [8], [9], but its other related applications were also successful [10]. One can observe that there is a straightforward analogy of a text and an image. In a text, the basic component is a letter, which composes words, and then words are organized into sentences, which form paragraphs, and so on. An image is composed of pixels, edges, lines, corners, polygons, objects, and these elements finally result in a scene. The words in our case are the lines and corners of the image. The polygon recognition method presented in the following sections is rather an experiment of combining the results of cognitive informatics and a leading edge algorithm for categorization. The goal of this paper is to make a further step towards object recognition with a cognitive approach to vision. The organization of this paper is as follows: In section II the visual feature array will be briefly described. Section III gives an introduction to UFEX. The implementation of the recognition system with two different representations of polygons is presented in section IV. This is followed by the evaluation of the proposed method and a comparison with other neural network based solutions in section V. Finally, section VI concludes the paper. II. THE VISUAL FEATURE ARRAY In this section the visual feature array (VFA) is presented, which is the basic component of the proposed method. VFA 1-4244-9713-4/06/$20.00 ©2006 IEEE 539