Bare Printed Circuit Board Defects Localization and Detection Using YOLOv5 Models

A printed circuit board (PCB) is a vital component of any electronic device. Over the years, the need to manufacture PCBs in large volumes has become a necessity due to technological advancement and the expansion of the consumer market. Small defects in the PCB signal trace can have adverse consequences on the functionality of an electronic device. Initially, the quality control process of PCB inspection was carried out manually which was not only tedious but also susceptible to human error. To automate this process, different techniques were developed using classical machine vision that comprised reference comparison, subtraction of images and morphological operations. The methods were effective and reduced inspection time but were sensitive to noise, rotation, and scaling of images. It was difficult to detect defects in complex patterns and only a small fraction of defects could be detected through classical techniques. In this study, four YOLOv5 models were tested for the classification and localization of defects in PCBs. Transfer learning on the bare PCB defect dataset was implemented using pretrained weights of YOLOv5. The PCB dataset was composed of 2,079 images with 6 classes of PCB defects. Data augmentations were carried out to increase the dataset which enhanced performance and generalizability of the models. The models were trained and tested at different image resolutions. Promising experimental results were achieved with the medium YOLOv5 model reaching a mean average precision of 96.92%.