Detection of Printed Circuit Board Defects on ENIG and ENIPIG Surface Finishes with Convolutional Neural Networks and Evaluation of Training Parameters

Increasingly high demands are being placed on the quality inspection of printed circuit boards (PCBs). A full surface inspection of all produced PCBs and a high defect detection accuracy of the inspection system are becoming prerequisites for an efficient quality management. At the same time, the demand for PCBs is constantly increasing over the years due to the high demand for electrical devices. Human inspection is no longer feasible due to the high production rates and required defect detection accuracy. Therefore, automatic inspection systems are increasingly used for quality control in the various process steps of PCB production. In this article, the first automatic inspection system for detecting defects on Electroless Nickel Immersion Gold (ENIG) and Electroless Nickel Immersion Palladium Immersion Gold (ENIPIG) surfaces is presented. A pretrained convolutional neural network (CNN) and the sliding window approach are used. A training dataset consisting of six different defect types and an OK class containing only defect-free PCB images was labeled for this classification problem. The hyperparameters learning rate and batch size are varied for different training runs of the CNN, and the performance of the network in PCB defect detection is evaluated using a test dataset. The true-positive rate, truenegative rate, and F1-score were analyzed for the evaluation. Our results show that the best performances could be achieved at low batch sizes and low learning rates.