Neural-based color image segmentation and classification using self-organizing maps

This paper presents a method for color image segmentation which uses classification to group pixels into regions. The chromaticity is used as data source for the method because it is normalized and considers only hue and saturation, excluding the luminance component. The classification is carried out by means of a self-organizing map (SOM), which is employed to obtain the main chromaticities present in the image. Then, each pixel is classified according to the identified classes. The number of classes is a priori unknown and the artificial neural network that implements the SOM is used to determine the main classes. The detection of the classes in the SOM is done by using a K-means segmentation. The obtained results substantiate the feasibility of the method, whose performance is compared, for evaluation, to human-assisted segmentation. A comparison of the method with a segmentation based on the k-nearest-neighbor classification is also presented.