Multi-view light source estimation for automated industrial quality control

Abstract

Nowadays, Computer Vision comprises a lot of different useful applications, ranging from real-time surveillance to series production. Since most of the mass-production systems are automatized, there is a non-zero probability for an error to happen, which in some cases would make the final product useless. In that regard, many factories often integrate into their assembly line a specialized camera system for the purpose of identifying errors at the end or during the production. In particular, a device composed of a network of calibrated cameras is usually employed to identify metrological and visual defects. Together with cameras, such device could comprise multiple light sources to easily spot defects caused by an imperfect structure of the object's surface. This work explores the possibility of increasing the precision of such an inspection system by estimating the position of each light source. The estimation is carried out through the analysis of the shading observed on the surface of some spherical Lambertian objects in different poses and from different points of view.