Detection and Classification of PCB Defects using Deep Learning Methods

Abstract

Automatic inspection methods are needed to cost-effectively discover defects in fabricated printed circuit boards (PCBs). This paper adapts a deep convolutional neural network design, called RetinaNet, to simultaneously locate and classify defects in PCBs. To improve detection performance, several network designs with regularization and dropout hyperparameters were tested. Not all defects render a PCB nonfunctional. This paper proposes re-evaluating network performance downweighing cosmetic defects. Adding a 25% dropout rate to each layer of the ResNet backbone, achieved an impressive 95.03% accuracy and 97.09% F-score on PCB images from a publicly available dataset. The study demonstrates the effectiveness of designing a single deep learning network to reliably and accurately detect as well as classify defects in PCB images. Our design can be extended to identification of defects in microelectronic circuits and semiconductor circuits with suitable modifications.