Defect Classification from Electronic Card Images by Deep Learning

Deep learning is a subclass of machine learning that puts the human brain-like learning process into practice. In this study, the performance of the Deep Learning method to detect selected manufacturing defects occurring during the production phase on the printed circuit boards (PCB) was examined and the results obtained by different training and design forms of the networks are discussed. In the frame of this study, short circuits, lifted components (tombstones), shifted components and general solder defects, which are very common as board defects, have been taken into account. The first design approach taken for the comparison is to train all data sets together with the combined training set and to provide the results. The second approach is a framework in which classes are separated and trained under a modular design. While the first method was successful in overall, solder defect results under modular training showed better performance above the combined training and this result supported the prediction of the use of a unified structure in the future work.