In this paper, we describe a new scheme to color image segmentation which is based on supervised pixel classification methods. Using color pixel classification alone does not extract accurately enough color regions, so we suggest to use a strategy based on four steps : simplification, pixel classification, marker extraction and color watershed growing. We detail on this paper the pixel classification and marker extraction steps. We present a quantitative measure which evaluate the resulting classifications and segmentations with a set of reference images. Our strategy is applicable to the detection of color objects in noisy environment and is particularly efficient on cytological color images.