In this paper, a new approach for the recognition and classification of convex objects in color images is presented. It is based on a collaboration between color quantization, mathematical morphology and reflectance estimation from RGB data. This yields a robust algorithm regarding the conditions of illumination, the color sensor used for acquisition, as well as the shape/overlapping ambiguities among the objects. One singularity of this work is the use of mathematical morphology in two distinct topologies: first in the entire image, for segmentation purposes, then locally, to enhance the classification of each object. A resolution reduction is used to alleviate the effect of local disturbances such as noise or natural impurities on the objects. The method's efficiency and usefulness are illustrated on the particular task of coffee beans sorting.