Printed Circuit Board Surface Defect Inspection via YOLOv5 Deep Learning Algorithm

Printed Circuit Boards (PCBs) surface defects produced during manufacturing process negatively impact the quality of products, which directly influences the stability and reliability of equipment performance. Nevertheless, it remains extremely difficult to detect small defects on the surface of PCB under the complicated background owing to its compact design. To address this problem, the version five of the state-of-the-art You Only Look Once (YOLO) algorithm, named YOLOv5, was utilized to detect different categories of PCB surface defects, namely Good, Moderate and Bad. To create the three categories of PCB surface defects, several resistors were soldered onto the surfaces of 30 PCBs, with each category of surface defect being illustrated by 10 PCBs, respectively. The number of images collected from all the soldered PCB surfaces for Good, Moderate and Bad PCB surface defect categories were 189, 158 and 175. Augmentation techniques, namely Auto-Adjust Contrast, 90-degree rotation clockwise and counterclockwise, shear, and ±25% exposure were executed on images collected to generate different images of three categories, generating 1400 images in total for training and testing of YOLOv5 model. Based on model results, the precision, recall and F1 score of the YOLOv5 model were 99.3%, 100% and 99.6%. The trained model obtained a 99.4% mean average precision (mAP) value with a high frames per second (FPS) value of 106.3 FPS for real-time detection and localization of different categories of PCB surface defects. These mAP and FPS values of the YOLOv5 model shows the model’s high performance in detecting PCB surface defect categories.