Collective classification for the detection of surface defects in automotive castings

Iron casting production is a very important industry that supplies critical products to other key sectors of the economy. For this reason, these castings are subject to very strict safety controls to ensure their final quality. One of the most common flaws is the appearance of defects on the surface. In particular, our work focuses on three of the most typical defects in iron foundries: inclusions, cold laps and misruns. We propose a new approach that detects these imperfections on the surface by means of a segmentation method that flags the potential defective regions on the casting and, then, applies collective classification techniques to determine whether the regions are defective or not. We show that these classifiers obtain high precision rates whilst decreasing the effort of labelling.