Acquisition and evaluation of illumination series for unsupervised defect detection
Analyzing scenes under variable illumination has been an important and widely studied research area in the field of machine vision. In this article, we present an illumination device for capturing image series of small objects under variable illumination directions. Due to using a digital projector as programmable light source and a parabolic reflector to reflect the emitted illumination patterns, the device dispenses with the need of moving parts. Furthermore, we demonstrate the utility of illumination series for unsupervised surface defect detection by applying statistical anomaly detection to the measured reflectance data. To this end, we show how relevant illumination directions can be determined without using labeled information by a clustering-based approach.