||The problem of assessing the quality of surfaces with confluent mottling are encountered e.g. in textile and metal industry. Such surfaces are characterized by an irregular appearance of spots, blotches or stains. Its quality assessment typically depends on the distribution of such phenomena, their sizes, densities, intensities, occurrences and spatial arrangement. The assessment also has to take topological properties like connectedness of such structures into account. The reason is that e.g. closely adjacent middle-sized spots might be considered of the same quality as a large connected spot, whereas a conglomeration of many middle or small-sized, but spatially distinctively separated spots with the same total mass has to be judged in a different way. The paper discusses two approaches for dealing with such quality inspection problems. We focus on problems for which a database of good and bad labeled sample images are available such that supervised machine learning techniques like support vector machines can be applied. The central question of the paper is how the classification rate of the learning algorithm under consideration can be improved by an appropriate feature extraction. For this the usefulness of maximally stable extremal region (MSER) method is investigated and compared with a straightforward approach based on standard threshold and blob analysis methods. Experimental studies of coated adhesive metal surfaces demonstrate that the approach with MSER for the feature generation leads to more robust results showing even better classification rates.