基于双重注意力机制优化的C3D视频火灾烟雾分类方法  

Video fire smoke classification method based on optimized C3D by dual attention mechanism

作　　者：朱家哲 何豪 阳书林 杨智 黄冬梅[1] ZHU Jiazhe;HE Hao;YANG Shulin;YANG Zhi;HUANG Dongmei(College of Energy Environment and Safety Engineering,China Jiliang University,Hangzhou 310018,China;Ningbo Siterwell Import and Export Co.,Ltd.,Ningbo 315000,China)

机构地区：[1]中国计量大学能源环境与安全工程学院,浙江杭州310018 [2]宁波赛特威尔进出口有限公司,浙江宁波315000

出　　处：《现代电子技术》2025年第5期53-58,共6页Modern Electronics Technique

基　　金：浙江省“尖兵”“领雁”研发攻关计划项目:工业企业安全生产智能防控关键技术-工业企业火灾灾变机理与感知预警处置一体化技术研究及应用(2024C03252);2024年度应急管理研发攻关科技项目:基于时空AI的火灾图像特征挖掘机类型辨识方法研究(2024YJ007)。

摘　　要：现有研究普遍针对特定类别的火灾烟雾制定模型算法来提高火灾识别的准确度,并没有对火灾进行精确分类。在发生火灾时,明确火灾类别对于后续火灾的扑灭与救援活动有指导作用。对此,文中开展了四种标准火实验,建立四种基本类型火灾(木材热解阴燃火、棉绳阴燃火、聚氨酯泡沫火、正庚烷油火)的视频图像数据集,并提出一种基于优化C3D卷积网络的视频火灾烟雾分类模型,为提升模型特征提取能力引入双重SE注意力模块,采用全局平均池化层(GAP)替代传统的全连接层,减少模型参数、防止过拟合,提升模型鲁棒性。实验结果表明,优化后的C3D模型在识别火灾烟雾类型方面准确率达到98.9%,相比原始模型准确率提升了9.28%,同时模型参数数量减少了64.39%,这为火灾烟雾监测与预警提供了重要的应用价值。The existing studies generally focus on specific categories of fire smoke and develop model algorithms to improve the accuracy of fire identification,but do not classify fires accurately.In the event of a fire,the clear fire category has a guiding effect on the subsequent fire suppression and rescue activities.In this paper,four standard fire experiments are carried out to establish video image data sets of four basic types of fires(wood pyrolysis smold fire,cotton rope smold fire,polyurethane foam fire,and n-heptane oil fire),and a video fire smoke classification model based on optimized C3D(Convolutional 3D)convolutional network is proposed.A dual SE(squeeze and excitation)attention module is introduced to improve the feature extraction capability of the model.A global average pooling(GAP)layer is adopted to replace the traditional full connection layer,which reduces model parameters,prevents overfitting,and improves the robustness of the model.The experimental results show that the accuracy rate of the optimized C3D model in identifying the types of fire smoke is 98.9%,which is 9.28%higher than that of the original model.In addition,the number of the model parameters is reduced by 64.39%.To sum up,the research can provide important application value for fire smoke monitoring and early warning.

关 键 词：深度学习 烟雾分类 C3D 注意力机制 火灾识别 准确度 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]