Aluminum Surface Defect Detection Method Based on DAS‐YOLO Network

Résumé fourni par la source

ABSTRACT To address the accuracy limitations of current methods in detecting aluminum surface defects, particularly those with small sizes and high variation, an aluminum surface defect detection algorithm named DAS‐YOLO, based on an improved YOLOv8n, is proposed. The C2f module in YOLOv8's backbone is enhanced by incorporating DCNv2, which improves the model's ability to handle irregular shapes and geometric transformations during feature extraction. An auxiliary training head (Aux Head) is added to capture multi‐scale and multi‐level features, significantly boosting small defect detection. Additionally, the traditional CIoU loss function is replaced with the Wise‐SIoU loss, accelerating convergence and enhancing both detection and regression accuracy. Experimental results on the Alibaba Tianchi aluminum surface defect dataset show that DAS‐YOLO achieves a mean average precision (mAP) of 85.3%. Compared to YOLOv8n, mAP50 improves by 3%, while precision and recall increase by 1.1% and 4.6%, respectively. Furthermore, to validate the model's performance on small defects and its generalization ability, it achieves a detection accuracy of 94.8% on the PCB dataset, with an mAP increase of 3.1% compared to YOLOv8n. These results demonstrate that DAS‐YOLO significantly enhances detection accuracy while maintaining speed and exhibits outstanding performance in small defect detection.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.