Deep Learning for Deflectometric Inspection of Specular Surfaces

Deflectometric techniques provide abundant information useful for aesthetic defect inspection in specular and glossy/shinny surfaces. A series of light patterns is observed indirectly through their reflection on the surface under inspection, and different geometrical or texture information about the surface can be extracted. In this paper, we present a deep learning based approach for the automated defect identification in deflectometric recordings. The proposed learning framework automatically learns features used for classification. Although the method is in an early stage of development, the experiments with industrial parts show promising results, and a very direct application if compared to hand-crafted feature definition approaches.