Abstract
Detecting defects in manufactured objects is essential for achieving high quality standards. To this purpose, automated visual inspection involves detecting defects analysing images captured by machine-vision systems. Nowadays, most of the methods depend on data-driven approaches. We propose a novel non-data-driven approach for detecting defects concerning the appearance of objects with arbitrary shapes, only requiring the CAD model and the knowledge of material distribution on the surface. We address the task with an inverse rendering approach, employing a differentiable physically-based renderer. We further estimate ambient-light, making the method able to deal with unknown light conditions in each inspection. Defects can be then localized from the generated heatmap computing the distance between the rendered images and the real ones. In this study, we examine a case involving steel cooling jackets for e-motors, transported by a conveyor belt in a production line. We leverage the motion of the object acquiring images in multiple time instants to increase the number of views. As there are no benchmarks available for comparisons, we provided only qualitative results on a real scenario. We offer a simple solution for visual inspection. It employs a machine-vision system exempt from lighting components, and it does not require data and training models unlike typical data-driven approaches.