摘要
存储石油化工产品时有极高的安全监测要求,传统油库监测系统采用人工观察视频监控的方式检测产品的存储是否存在安全问题,该方法在火灾应急行动中存在时效性低、响应慢及漏检等问题。针对上述问题,文章提出了一种基于深度学习的油库安全监测方法,在原始YOLOv5(单阶段目标检测算法)的主干网络添加坐标注意力机制(Coordinate Attention,CA),该机制可以得到方向感知和位置感知的注意力图,将这些图应用于输入特征图,可以更加丰富地表示感兴趣对象;使用回归损失函数EIOU_Loss代替原有的GIOU_Loss进行预测框的定位回归损失计算,该方法能提高Bounding Box(边界框)的回归精度;使用DIOU-NMS替换NMS(非极大抑制),可以提高对遮挡目标的辨识度。该方法在自制的测试数据集中的mAP达到了91.2%,说明该方法在油库安全监测领域具有较高的应用价值。
Petrochemical products storage requires extremely high level of safety monitoring. Traditional oil depot monitoring system uses manual observation video monitoring to detect whether there are safety problems. This method has problems such as low timeliness, slow response and missed detection. In response to these problems, this paper proposes an oil depot safety monitoring method based on deep learning. The Coordinate Attention(CA) mechanism is added to the backbone network of original YOLOv5(a single-stage target detection algorithm), so as to obtain the attention maps of direction awareness and position awareness. These maps are applied to input feature maps so that the objects of interest are represented more abundantly. Then, the original GIOU_Loss is replaced by EIOU_Loss function to calculate the localization regression loss of the prediction box, which can improve the regression accuracy of Bounding Box. Finally, DIOU-NMS is used to replace NMS(Non-Maximum Suppression) to improve the recognition of occluded targets. The mAP of this method in the self-made test dataset reaches 91.2%, indicating that the proposed method has high application value in the safety monitoring of oil depots.
作者
林海
胡旭晓
吴跃成
汪燕超
LIN Hai;HU Xuxiao;WU Yuecheng;WANG Yanchao(School of Mechanical Engineering and Automation,Zhejiang Sci-Tech University,Hangzhou 310018,China)
出处
《软件工程》
2023年第3期15-17,14,共4页
Software Engineering
