摘要
针对传统的油田井场的灭火器、过桥盖板、压力表等安全设施中实时目标检测算法精确度低、实时性差的问题,将井场设施安全性的跟踪监测作为研究对象,提出了一种基于改进YOLOv8的井场设施安全实时监测算法DCH-YOLO。首先,为了更好地适应复杂场景,在YOLOv8骨干网络的C2f模块中添加可变形卷积,提升目标的定位能力和模型自适应能力,以改善YOLOv8对遮挡或旋转后的目标及目标尺度变化较大时出现误检或漏检的现象。其次,引入CBAM注意力机制,以提高网络对重要图像区域的关注度,增强空间注意力,提升网络的泛化能力。最后,采用边界框回归损失函数HIoU取代CIoU,能够动态调整边界框回归损失,提高目标监测的准确性。实验结果表明,DCH-YOLO检测精确率较原始YOLOv8模型提升了4.7%,mAP50和mAP50:95分别提升了5.43%和4.2%,对现场使用具有一定的参考价值。
Aiming at the problems of low accuracy and poor real-time performance of traditional real-time target detection algorithms of safety facilities such as fire extinguisher,bridge cover plate,pressure gauge,etc.,this paper takes the tracking monitoring of safety of well site facilities as the research object,and proposes a real-time monitoring algorithm DCH-YOLO based on improved YOLOv8 for safety of well site facilities.First of all,in order to better adapt to complex scenes,deformable convolution is added to the C2f module of YOLOv8 backbone network to improve the target positioning capability and model adaptive capability,so as to improve the false detection or missed detection of YOLOv8 when the target is occluded or rotated and the target scale changes greatly.Secondly,the CBAM attention mechanism is introduced to improve the network's attention to important image regions,enhance spatial attention,and improve the generalization ability of the network.Finally,the boundary regression loss function HloU is used to replace CloU,which can dynamically adjust the boundary regression loss and improve the accuracy of target monitoring.The experimental results show that the detection accuracy of DCH-YOLO is 4.7%higher than that of the original YOLOv8 model,and mAP50 and MAP50:95 are 5.43%and 4.2%higher,respectively,which has certain reference value for field use.
作者
程诗蕾
程国建
CHENG Shilei;CHENG Guojian(School of Computer Science,Xi'an Shiyou University,Xi'an,Shaanxi 710065,China)
出处
《石油工业技术监督》
2024年第9期45-50,共6页
Technology Supervision in Petroleum Industry
关键词
井场设施
YOLOv8
实时监测
回归损失
well site facilities
YOLOv8
real-time monitoring
regression loss
作者简介
程诗蕾(2000-),女,硕士研究生,主要研究方向为深度学习中的目标检测及识别技术。E-mail:15991658155@163.com。
