Tiny Surface Defects on Small Ring Parts Using Normal Maps

Detection of tiny surface defects on small ring parts remains challenging due to the unnoticeable visual features of such defects and the interference of small surface scratches. This paper proposes a novel method for detecting tiny surface defects based on normal maps of metal parts. To better characterize features of tiny defects and differentiate them from small scratches, we recover the normal map of the metal part through analyzing its directional reflections obtained with our specifically designed directional light units. Based on the normal map, a cascaded detector trained by the AdaBoost approach combined with the joint features and fast feature pyramid is used to localize the defects, achieving fast and accurate detection of tiny surface defects. The proposed method can achieve high detection accuracy with extremely fast speed, only 23 ms per metal part, and comparisons against other methods show our superiority.