Color image segmentation is a rapidly developing technique of digital image processing. The goal of image segmentation is to divide the image into non-overlapping homogeneous regions or objects. Object extraction, object recognition, object-based compression, and content-based image retrieval are typical applications. Recently, a large number of techniques and algorithms have been proposed for image segmentation. Among them, those based on watershed transformation can be effectively used for segmentation of grey-scale and color images. However, conventional watershed transform may produce a severe over-segmentation of the image, i.e., many small regions are produced due to many local minima in the input image. In this paper we propose an algorithm for eliminating irrelevant minima in the resulting gradient images. We use gradient-threshold-controlled segmentation function for watershed transformation. The global gradient threshold is obtained from the first derivative of the region-function with respect to threshold. The experimental results show that the proposed algorithm can effectively improve segmentation accuracy for different image natures, such as presence or not of non-homogeneous illumination, texture, contours, and shadows.
