Seminar on Intelligent Technology and Its Applications 2009 ISSN 2085 – 9732 A New Color Segmentation Method Based on Normalized RGB Chromaticity Diagram Aryuanto 1) Komang Somawirata 2) F. Yudi Limpraptono 3) Department of Electrical Engineering, Faculty of Industrial Technology ITN Malang, Indonesia. 1) Email: aryuanto@fti.itn.ac.id Abstract - This paper presents the new color segmentation method based on normalized RGB chromaticity diagram by utilizing the line in the chromaticity diagram to separate colors. Parameter of the line is obtained automatically using the peak and valley analysis of the newly developed histograms. The method is simple, fast, and effective to overcome the problem of illumination changes. The method shows the promising results for color segmentation of the outdoor sign images. Keywords: color segmentation, normalized RGB, chromaticity diagram, sign recognition. 1. INTRODUCTION Image segmentation plays an important role in the computer vision fields. It is usually used as the preliminary process for high level image processing. It extracts the meaningful object or object of interest from a whole image. Basically, image segmentation is to partition an image into non-overlapping regions [1]. A region is a homogenous group of connected pixels having a particular property, such as color, gray level, texture, motion, etc. In the color images, color segmentation is very popular method, since color is more meaningful property to the human perception. Hence, the task of color segmentation is to separate objects according to the colors. Once objects are separated by color segmentation, further process such as object recognition could be performed to extract the meaningful information from the image. Many applications employ the color segmentation for understanding an image, such as for interactive robot [2,3,4], video retrieval [5], road sign recognition [6,7,8]. In outdoor environments, the illumination changes could not be controlled, hence color segmentation should be robust to this problem. Furthermore, the algorithm shoud be fast enough for real-time implementation. Due to the facts, color segmentation still becomes a challenging topic and offers the open area for improvement. The existing color segmentation methods could be classified into several approaches : histogram-based method, boundary-based method, region-based method, Neural network-based method, Fuzzy-based method, and Genetic Algorithm (GA)-based method. Histogram-based method is commonly used for monochrome image segmentation. Since color images are usually represented by three dimensional (3D) of RGB colors, the histogram-based color segmentation basically fuses the three thresholds obtained from each color channel. In [5], they employed a gray-level thresholding method in each of the color Red (R), Green (G), Blue (B), then used the generated threshold values as a base to produce a set of desired multiple threshold values for video image segmentation by means of an unsupervised clustering process. In the boundary-based method, an edge detector is employed to find the boundary of an object. This method works by the fact that the the pixel intenstity values will change rapidly at the boundary of two regions. For color segmentation, at first edge detection is performed to each color channel R,G,B separately. Then the resulted edges are merged to obtain the final edge image. In the region-based method, pixels are grouped according to the homogenity criteria. The region growing, and split and merge algorithms are the examples of the region-based method. In the region growing algorithm, pixels or subregions are grouped into larger regions based on predefined criteria [9]. The algorithm starts with a set of seed points and then grows regions by appending to each seed those neighboring pixels that have similar properties to the seed, such as gray level or color. On the contrary, the split and merge algorithm subdivide an image initially into a set of arbitrary, disjointed regions and then merge and/or split the regions in attemp to satisfy the predefined criteria [9]. These techniques have two main drawbacks [1]: They are both strongly dependent on global predefined criteria; while the region growing technique depends also on initial segments, which is the first pixel/segment to be scanned, and the order of the process. The Artificial neural network that implements self- organizing map (SOM) is used for color segmentation as proposed by [10]. They use normalized RGB chromaticity as the data source of the neural network. During learning phase, sample points are taken from the image and submitted to the network. When the feature map has been formed, the main chromaticities