基于YOLOv5的火灾识别模型设计  

Fire Detection Model Design Based on YOLOv5

作　　者：唐昉[1] 马永智 王书恒 沈文韬 刘铭骥 夏子潮 TANG Fang;MA Yong-zhi;WANG shu-heng;SHEN Wen-tao;LIU Ming-ji;XIA Zi-chao(Wuhan Second Ship Design and Research Institute,Wuhan 430064,China;Marine Design and Research Institute of China,Shanghai 200011,China;Institute of Advanced Technology,China University of Geosciences(Wuhan),Wuhan 430074,China;Power Engineering College,Naval University of Engineering,Wuhan 430033,China)

机构地区：[1]武汉第二船舶设计研究所,武汉430064 [2]中国船舶及海洋工程设计研究院,上海200011 [3]中国地质大学(武汉)先进技术研究院,武汉430074 [4]海军工程大学动力工程学院,武汉430033

出　　处：《船海工程》2025年第2期86-92,共7页Ship & Ocean Engineering

基　　金：海军工程大学自主研发计划(2023502060);中央高校基本科研业务费专项资金(162301222607,162301212668)。

摘　　要：针对深度学习方法中存在的特征提取能力不足和小目标检测能力较弱问题,提出一种基于改进YOLOv5神经网络模型的火灾识别方法。为提高火焰和烟雾的检测精度并降低误检率,通过在主干网络中添加注意力机制改进YOLOv5的主干网络,提高模型对图像特征的学习能力。采用新的特征融合方式取代原网络的特征信息融合方式,改进损失函数优化先验框的回归定位精度。通过将颈部网络部分FPN结构替换为BiFPN结构,并使用EIOU损失函数,提升模型对小目标的识别能力和识别精度。To address the insufficient feature extraction capability and weak small target detection ability in deep learning methods,a fire detection method was proposed based on an improved YOLOv5 neural network model.To enhance the detection accuracy of flames and smoke while reducing the false detection rate,the backbone network of YOLOv5 was improved by adding attention mechanisms,which enhances the model’s ability to learn image features.A new feature fusion method replaced the original network’s feature information fusion method,and the loss function was improved to optimize the regression localization accuracy of anchor boxes.By replacing the neck network’s FPN structure with a BiFPN structure and using the EIOU loss function,the model’s ability to recognize small targets and its recognition accuracy were improved.

关 键 词：火灾识别 火焰检测 注意力机制 YOLOv5 

分 类 号：U662[交通运输工程—船舶及航道工程]