Precise Detection for Dense PCB Components Based on Modified YOLOv8

• Published in: IEEE access Vol. 11; p. 1
• Main Authors: Ling, Qin, Isa, Nor Ashidi Mat, Asaari, Mohd Shahrimie Mohd
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.01.2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Artificial neural networks; Circuit boards; Computational modeling; Convolution; Costs; Feature extraction; Frames per second; Ghost convolution; High-precision; Image processing; Inspection; Lightweight; Machine learning algorithms; Optimization; PCB component detection; Precision engineering; Printed circuits; Sig-IoU loss
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2023.3325885

Effective detection of dense printed circuit board (PCB) components contributes to the optimization of automatic flow of production. In addition, PCB component recognition is also the essential prerequisite for early defect detection. Current PCB component detection approaches are not adept in both rapid and precise detection. YOLOv8 models have exhibited effective performances for detecting common objects, such as person, car, chair, dog etc. However, it is still tricky for YOLOv8 models to inspect dense and disparate PCB components precisely. Thus, a novel convolution neural network (CNN) model is proposed for dense PCB component detection by introducing several modifications onto YOLOv8. First, creative C2Focal module is designed as the core element of the backbone, combining both fine-grained local and coarse-grained global features concurrently. Then, the lightweight Ghost convolutions are inserted to effectively reduce the computation cost, meanwhile maintaining the detection performance. Finally, a new bounding box regression loss that is Sig-IoU loss, is proposed to facilitate the prediction regression and promote the positioning accuracy. The experiments on our PCB component dataset demonstrate that our proposed model performs the highest mean average precisions of 87.7% (mAP@0.5) and 75.3% (mAP@0.5:0.95) respectively, exceeding other state-of-the-arts. Besides, the detection speed hits 110 frames per second using RTX A4000, which is potential to realize the real-time inspection.