Wei Ze, Yang Fan, Zhong Kezhen, Yao Linkun
School of Electronic and Information Engineering, Hebei University of Technology, Tianjin, China.
Faculty of Computing, Harbin Institute of Technology, Harbin, China.
PLoS One. 2025 Jun 2;20(6):e0323684. doi: 10.1371/journal.pone.0323684. eCollection 2025.
Nowadays, industrial electronic products are integrated into all aspects of life, with PCB quality playing a decisive role in their performance. Ensuring PCB factory quality is thus crucial. Common PCB defects serve as key references for evaluating quality. To address low detection accuracy and the bulky size of existing models, we propose an improved PCB-YOLO model based on YOLOv8n.To reduce model size, we introduce a novel CRSCC module combining SCConv convolution and C2f, enhancing PCB defect detection accuracy and significantly reducing model parameters. For feature fusion, we propose the FFCA attention module, designed to handle PCB surface defect characteristics by fusing multi-scale local features. This improves spatial dependency capture, detail attention, feature resolution, and detection accuracy. Additionally, the WIPIoU loss function is developed to calculate IoU using auxiliary boundaries and address low-quality data, improving small-target recognition and accelerating convergence. Experimental results demonstrate significant improvements in PCB defect detection, with mAP50 increasing by 5.7%, and reductions of 13.3% and 14.8% in model parameters and computational complexity, respectively. Compared to mainstream models, PCB-YOLO achieves the best overall performance. The model's effectiveness and generalization are further validated on the NEU-DET steel surface defect dataset, achieving excellent results. The PCB-YOLO model offers a practical, efficient solution for PCB and steel defect detection, with broad application prospects.
如今，工业电子产品已融入生活的方方面面，印刷电路板（PCB）的质量对其性能起着决定性作用。因此，确保PCB工厂的质量至关重要。常见的PCB缺陷是评估质量的关键参考。为了解决现有模型检测精度低和体积庞大的问题，我们提出了一种基于YOLOv8n的改进型PCB-YOLO模型。为了减小模型尺寸，我们引入了一种结合SCConv卷积和C2f的新型CRSCC模块，提高了PCB缺陷检测精度，并显著减少了模型参数。对于特征融合，我们提出了FFCA注意力模块，旨在通过融合多尺度局部特征来处理PCB表面缺陷特征。这提高了空间依赖性捕捉、细节注意力、特征分辨率和检测精度。此外，还开发了WIPIoU损失函数，利用辅助边界计算交并比（IoU），解决低质量数据问题，提高小目标识别能力并加速收敛。实验结果表明，PCB缺陷检测有显著改进，mAP50提高了5.7%，模型参数和计算复杂度分别降低了13.3%和14.8%。与主流模型相比，PCB-YOLO实现了最佳的整体性能。该模型在NEU-DET钢表面缺陷数据集上进一步验证了有效性和泛化能力，取得了优异结果。PCB-YOLO模型为PCB和钢缺陷检测提供了一种实用、高效的解决方案，具有广阔的应用前景。
