Author:
Wang Luyang,Zhang Gongxue,Wang Weijun,Chen Jinyuan,Jiang Xuyao,Yuan Hai,Huang Zucheng
Abstract
In industrial aluminum sheet surface defect detection, false detection, missed detection, and low efficiency are prevalent challenges. Therefore, this paper introduces an improved YOLOv8 algorithm to address these issues. Specifically, the C2f-DSConv module incorporated enhances the network’s feature extraction capabilities, and a small target detection layer (160 × 160) improves the recognition of small targets. Besides, the DyHead dynamic detection head augments target representation, and MPDIoU replaces the regression loss function to refine detection accuracy. The improved algorithm is named YOLOv8n-DSDM, with experimental evaluations on an industrial aluminum sheet surface defect dataset demonstrating its effectiveness. YOLOv8n-DSDM achieves an average mean average precision (mAP50%) of 94.7%, demonstrating a 3.5% improvement over the original YOLOv8n. With a single-frame detection time of 2.5 ms and a parameter count of 3.77 M, YOLOv8n-DSDM meets the real-time detection requirements for industrial applications.
Reference34 articles.
1. YOLO9000: better, faster, stronger;Redmon,2017
2. SSD: single shot MultiBox detector;Liu,2016
3. Rich feature hierarchies for accurate object detection and semantic segmentation;Girshick;CoRR, abs,2013
4. Fast r-cnn;Girshick,2015
5. Faster r-cnn: towards real-time object detection with region proposal networks;Ren;Adv Neural Inf Process Syst,2015
Cited by
1 articles.
订阅此论文施引文献
订阅此论文施引文献，注册后可以免费订阅5篇论文的施引文献，订阅后可以查看论文全部施引文献