螺栓是电磁能装备中重要的连接紧固件,基于计算机视觉的检测有助于后续的螺栓松动检测、缺陷检测等.针对螺栓目标小的特征,提出了基于改进YOLOv5(you only look once)的螺栓目标检测方法.首先,对Backbone引入SE(squeeze-and-excitation)注意力机制模块,提高基础主干网络的特征提取能力;然后,引入Bi-FPN(bi-directional feature pyramid network)结构实现Neck 部分的特征融合功能,加权融合不同层的特征信息;最后,建立由个人收集和公共数据集NPU-BOLD组成的螺栓数据集,共包含 1 520 张图像.该数据集下的实验结果表明:该方法检测的mAP 为 90.0%,较YOLOv5 s 提升了 2.4%,精度提升较明显且能够保持较快的检测速度.
Bolt fastener detection based on improved YOLOv5
Bolts are important connecting fasteners in electromagnetic energy equipment.Computer vision-based detection is helpful for subsequent bolt loosening detection,defect detection and other steps.Aiming at the characteristics of small bolt target,a bolt detection method based on improved YOLOv5(you only look once)was proposed.Firstly,the SE(squeeze-and-excitation)attention mechanism module was introduced for Backbone to improve the feature extraction ability of Backbone networks.Then Bi-FPN(bi-directional feature pyramid network)structure was introduced to realize the feature fusion function of Neck part,and the feature information of different layers was weighted and fused.Finally,a bolt dataset consisting of personal collection and public dataset NPU-BOLD was estab-lished,which contained 1 520 images in total.The experimental results on this dataset show that the mAP detected by that method is 90.0%,namely 2.4%higher than that of YOLOv5s,and the accuracy is significantly improved with faster speed.
NETL
NSTL
万方数据
