Publication: Enhanced you only look once networks for detection of printed circuit board defects and components
Date
2024-07
Authors
Qin Ling
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Printed circuit boards (PCBs) are becoming increasingly complicated, diminutive and delicate due to the rapid development of integrated circuit technology.Effective detection for PCB defects and components is critically important andchallenging for the PCB industry. However, current methods for PCB inspection are hardly competent for both rapid and accurate detection simultaneously. They always achieve precise detection by introducing computationally expensive operations, which are adverse to their detection speeds. Even for You Only Look Once (YOLO) networks,
which are renowned for their accurate and real-time performance in object detection, it is still difficult to detect small PCB defects and components, because of the small dimensions, large scale variance, dense distribution and diverse appearances. Therefore, based on various modifications on YOLOv5, YOLOv7 and YOLOv8, three novel deep learning models, such as TD-YOLO (YOLO for tiny defect detection), SDD-Net (soldering defect detection network) and DC-YOLO (YOLO for dense component detection) are proposed respectively for rapid and precise PCB inspection. The improvements in TD-YOLO involve recomposed data augmentation, novel anchors design, the introduction of ShuffleNet block and an efficient feature pyramid network. For SDD-Net, a novel spatial pyramid pooling block, a hybrid combination attention mechanism, a residual feature pyramid network and an efficient intersection over union (IoU) loss function are implemented. For DC-YOLO, the modifications contain introducing Ghost convolution and novel C2Focal modules into the backbone, and a Sig-IoU loss. Consequently, TD-YOLO achieves outstanding mean average
precision (mAP) of 99.5% and the fastest speed of 37 frames per second (FPS) for PCB cosmetic defects in high-resolution images. SDD-Net attains the highest mAP of 99.1% with the 102 FPS speed for PCB soldering defects. DC-YOLO obtains the highest mAP of 87.7% and speed of 110 FPS for PCB components. Overall, DC- YOLO is the best method of the three proposed models in terms of the detection precision, because it not only exhibits excellent results for PCB components, but also has impressive generalization ability on both PCB cosmetic defects and soldering defects.