Automatic Edge Detection by Combining Kohonen SOM and the Canny Operator P. Sampaziotis and N. Papamarkos Image Processing and Multimedia Laboratory, Department of Electrical & Computer Engineering, Democritus University of Thrace, 67100 Xanthi, Greece papamark@ee.duth.gr Abstract. In this paper a new method for edge detection in grayscale images is presented. It is based on the use of the Kohonen self-organizing map (SOM) neural network combined with the methodology of Canny edge detector. Gradient information obtained from different masks and at different smoothing scales is classified in three classes (Edge, Non Edge and Fuzzy Edge) using an hierarchical Kohonen network. Using the three classes obtained, the final stage of hysterisis thresholding is performed in a fully automatic way. The proposed technique is extensively tested with success. 1 Introduction Changes or discontinuities in an image amplitude attribute such as intensity are fundamentally important primitive characteristics of an image because they often provide an indication of the physical extent of objects within the image. The detection of these changes or discontinuities is a fundamental operation in computer vision with numerous approaches to it. Marr and Hildreth [3] introduced the theory of edge detection and described a method for determining the edges using the zero-crossings of the Laplacian of Gaussian of an image. Canny determined edges by an optimization process [1] and proposed an approximation to the optimal detector as the maxima of gradient magnitude of a Gaussian-smoothed image. Lily Rui Liang and Carl G. Looney proposed a fuzzy classifier [2] that detects classes of image pixels cor- responding to gray level variation in the various directions. A fuzzy reasoning approach was proposed by Todd Law and Hidenori Itoh [8], in which image fil- tering, edge detection and edge tracing are completely based on fuzzy rules. The use of self-organising map and the Peano scan for edge detection in multispec- tral images was proposed by P.J. Toivanen and J. Ansamaki [5]. In [10], Pihno used a feed-forward artificial neural of perceptron-like units and trained it with a synthetic image formed of concentric rings with different gray levels. Weller [11] trained a neural net by reference to a small training set, so that a Sobel operator was simulated. In Bezdek’s approach [12], a neural net is trained on M. Lazo and A. Sanfeliu (Eds.): CIARP 2005, LNCS 3773, pp. 954–965, 2005. c  Springer-Verlag Berlin Heidelberg 2005