YOLO-DefXpert: An Advanced Defect Detection on PCB Surfaces Using Improved YOLOv11 Algorithm
<p dir="ltr">Surface defects in Printed Circuit Boards (PCBs), which arise during manufacturing, significantly impact product quality and directly influence equipment performance, stability and reliability. Accurately identifying small defects on PCB surfaces remains a considerable c...
محفوظ في:
| المؤلف الرئيسي: | |
|---|---|
| مؤلفون آخرون: | , , , , |
| منشور في: |
2025
|
| الموضوعات: | |
| الوسوم: |
إضافة وسم
لا توجد وسوم, كن أول من يضع وسما على هذه التسجيلة!
|
| الملخص: | <p dir="ltr">Surface defects in Printed Circuit Boards (PCBs), which arise during manufacturing, significantly impact product quality and directly influence equipment performance, stability and reliability. Accurately identifying small defects on PCB surfaces remains a considerable challenge, particularly under complex background conditions, due to the intricate and compact layout of the boards. This study introduces an improved PCB defect detection model, YOLO-DefXpert, using the YOLOv11 algorithm to address the low accuracy and efficiency challenges in detecting tiny-sized defects on PCBs. First, the standard backbone network of the YOLOv11 algorithm is replaced with the Swin Transformer to extract more robust features of defects, and the Convolutional Block Attention Module (CBAM) is added in the Patch Merging modules to alleviate feature leakage during the downsampling operation. Second, the standard convolutional operations are replaced with Deformable Convolutional Networksv2 (DCNv2) in the neck section to improve robustness in identifying multi-scale defects. Finally, an Additional Feature Fusion Layer (AFFL) is introduced in the neck to enhance the performance of the small defect identification. The effectiveness of the proposed YOLO-DefXpert is validated through experimental results obtained from publicly available PCB datasets. The proposed model achieves a mAP50 of 99.0% and a mAP95 of 60.6% in the HRIPCB benchmark dataset and 99.3% in mAP50 and 63.4% in mAP95 in the PCB dataset. Compared to the standard YOLOv11 model, the proposed YOLO-DefXpert attained an improvement of 9.3% and 13.2% in mAP50 and mAP95, an 11.25% increase in frames per second, and a 69.85MB decrease in model size. These findings highlight a notable enhancement in both accuracy and model efficiency in detecting tiny defects in the PCB board.</p><h2 dir="ltr">Other Information</h2><p dir="ltr">Published in: IEEE Access<br>License: <a href="https://creativecommons.org/licenses/by/4.0/deed.en" target="_blank">https://creativecommons.org/licenses/by/4.0/</a><br>See article on publisher's website: <a href="https://dx.doi.org/10.1109/access.2025.3595048" target="_blank">https://dx.doi.org/10.1109/access.2025.3595048</a></p> |
|---|