YOLOv5-MDS: Target Detection Model for PCB Defect Inspection Based on YOLOv5 Integrated With Mamba Architecture

• Published in: IEEE access Vol. 13; pp. 136612 - 136624
• Main Authors: Guo, Deming, Xi, Qi, Zhang, Long, Wang, Jing, Ye, Yilong, Zeng, Ziyi
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Artificial neural networks; Circuit boards; Computational modeling; Convolutional neural networks; Defect detection; Defects; Feature extraction; Image segmentation; Inspection; mamba; Object recognition; printed circuit board; Printed circuits; Production; Real-time systems; Target detection; tiny object detection; YOLO; YOLOv5
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2025.3591987

The Printed Circuit Board (PCB), which serves as the foundational component of numerous electronic devices, exhibits a complex relationship between its quality and the lifespan and performance of those products. However, annual losses are significantly contributed to by PCB defects arising from various production factors. PCB defects present significant challenges in detection, due to their small size and irregular distribution. Existing methods struggle to effectively detect these minute and scattered defects, especially in the context of industrial production, where high accuracy and speed are paramount. To address this issue, YOLOv5-MDS has been proposed, a convolutional neural network (CNN) specifically designed for image segmentation and classification. This model is derived from and enhanced on the YOLOv5 model. Our primary objective was to refine the neck section of the YOLOv5 model by integrating the Mamba architecture within it. Additionally, we optimized the intersection over union (IoU) calculation method and the upsampling module. This strategic integration aims to overcome the inherent limitation of convolutional operations' local receptive fields, thereby enabling the model to capture features more effectively. Our proposed improvements enhance detection accuracy while maintaining computational efficiency. Experimental results demonstrate that our model achieves significant improvements in accuracy for PCB defect detection tasks compared to the baseline model. Compared to the baseline model, our proposed model achieves a 34.3% performance improvement. When benchmarked on edge devices, YOLOv5-MDS demonstrates superior detection capabilities with mAP95 values exceeding YOLOv8n by 11.0% and YOLOv11n by 0.9%, while simultaneously showing significant efficiency advantages - operating 29% faster than YOLOv8n and 57% faster than YOLOv11n in inference speed. These comparative results comprehensively validate the practical applicability of YOLOv5-MDS in industrial deployment scenarios.