目的:针对钢面缺陷检测中特征提取不足或丢失的问题进行改进，从而提升检测的准确性。方法:提出一种基于注意力机制的DA-YOLO缺陷检测算法。该算法以YOLOv8s为基础网络，在主干网络末端加入一种新的双重注意力机制，以增强特征提取能力，避免缺陷信息丢失过多。同时，为了精确定位缺陷，引入MPDIoU损失代替CIoU进行边界框优化。结果:DA-YOLO算法在NEU-DET钢面数据集上平均精度均值(mean average precision，mAP)为 80。5％，比YOLOv8s增长了3。8％。结论:DA-YOLO 算法实现了更好的检测效果，相比YOLOv8s更适合应用于钢面缺陷检测任务中。
A DA-YOLO defect detection algorithm based on the attention mechanism
Aims:This paper raises the accuracy of detection results by improving insufficient extraction or loss of defect information for steel surface defects detection.Methods:A DA-YOLO defect detection algorithm based on the attention mechanism was proposed.Based on the YOLOv8s network,the algorithm added a new double attention mechanism at the end of the backbone network to enhance feature extraction ability and avoid excessive loss of defect information.At the same time,in order to accurately locate defects,MPDIoU loss was introduced instead of CIoU for bounding box optimization.Results:The mean average precision(mAP)of the DA-YOLO algorithm on the NEU-DET steel surface dataset was 80.5% ,which was an increase of 3.8% compared with YOLOv8s.Conclusions:The DA-YOLO algorithm achieves better detection results and is more suitable for steel surface defect detection tasks than the YOLOv8s algorithm.
国家科技期刊平台
NSTL
万方数据
