Detection of Printed Circuit Board Defects with Photometric Stereo and Convolutional Neural Networks

Abstract

The quality inspection of printed circuit boards (PCBs) is no longer feasible by human inspectors due to accuracy requirements and the processing volume. Automated optical inspection systems must be specifically designed to meet the various inspection requirements. A photometric stereo setup is suitable for the inspection of highly reflective gold areas on PCBs. In this process, several image acquisitions are performed under different illumination directions. This can reveal defects that are not visible under other illumination systems. In this paper, we use a segmented ring light so that image acquisition is possible under four different illumination directions. Using these images, a normal map and a mean image are calculated. The defects on the gold areas of PCBs are detectable in either the normal map, the mean image, or both images. A CNN for classification detects the defects. The input is a 6-dimensional image from normal map and mean image. An accuracy up to 95% can be achieved with the available dataset.