基于改进YOLOv8s的钢材表面缺陷检测算法

梁礼明,龙鹏威,金家新,李仁杰,曾璐

Steel surface defect detection algorithm based on improved YOLOv8s

Liming LIANG,Pengwei LONG,Jiaxin JIN,Renjie LI,Lu ZENG

表 7 不同算法在NEU-DET和Severstal数据集上的对比实验结果

Tab.7 Comparative experimental results of different algorithms on NEU-DET and Severstal datasets

数据集 模型方法 mAP/% Params/106 FLOPs/109 FPS/帧 NEU-DET Faster R-CNN 65.7 72.0 167.3 17 SSD 61.0 41.1 145.3 41 YOLOv3 67.0 61.5 155.0 31 YOLOv3-tiny 46.5 8.6 12.9 142 YOLOv4 51.0 52.5 119.8 45 YOLOv4-tiny 54.6 5.9 16.1 128 YOLOv5s 70.1 7.07 16.4 102 YOLOX-s 71.8 8.0 21.6 46 YOLOv7 70.0 37.2 104.8 36 YOLOv7-tiny 68.7 6.02 13.1 108 YOLOv8s 72.8 11.1 28.8 120 文献[11] 78.5 5.8 10.9 49 文献[18] 74.1 23.9 — 75 SDB-YOLOv8s(本研究) 79.2 7.2 16.2 146 Severstal SSD 65.3 41.1 145.3 12 YOLOv3-tiny 56.4 8.6 12.9 117 YOLOv4-tiny 59.6 5.9 16.1 103 YOLOv7-tiny 68.7 6.02 13.1 108 YOLOv5s 72.4 7.07 16.4 59 YOLOv5m 73.2 21.0 50.3 52.6 YOLOX-s 73.8 8.0 21.6 42 YOLOv8s 69.8 11.1 28.8 103 SDB-YOLOv8s(本研究) 76.9 7.2 16.2 121