Research on PCB and chip-wound inductor defect detection method based on computer vision

Abstract

Printed circuit boards (PCBs) are widely used in the electronics industry, typically for connecting electronic circuits. In electronics manufacturing, PCB quality control is crucial for product reliability. Traditional manual inspection methods are inefficient, costly, and prone to errors. Traditional machine vision technologies rely on manually designed features, resulting in poor robustness. With the development of deep learning networks, the YOLO series has become mainstream, but it still is not perfect in PCB and inductor defect detection. This thesis proposes an efficient and lightweight defect detection algorithm based on YOLOv8n. To optimize the network architecture, FasterNet is used to replace the original YOLO network in the backbone, significantly reducing the number of parameters. Simultaneously, an AFPN is used in the head, and the number of output channels is compressed from 256 to 128. Finally, experiments show that the improved model performs excellently, reducing the number of parameters from 3.01 million in YOLOv8n to 2.43 million. In PCB defect detection, the Precision reaches 97.6%. In inductor defect detection, the model's inference time for a single image is only 2.45 milliseconds.