摘要
我国火灾隐患高居不下,严重威胁着人民群众的生产、生活安全,火灾释放的污染物对空气质量和人体健康都有潜在影响,提高视频火灾烟雾识别技术能够有效地把火灾扼杀在摇篮里。本文提出了一种基于多尺度特征融合的烟雾识别模型,在神经网络模型VGG19的基础上构建了自上而下的特征路径和自下而上的注意路径,在进行卷积的同时保留低层特征,更利于烟雾识别这项任务。本研究采用火灾科学国家重点实验室公开的烟雾图像数据集进行实验,准确率达到99.75%,相较于仅使用神经网络模型VGG19提高了9.85%。
Hidden dangers of fire are high in our country,seriously threatening the safety of production and life of people.Pollutants released by fire have potential effects on air quality and human health.Improving video fire smoke identification technology can effectively kill fire in the cradle.In this paper,a smoke recognition model based on multi-scale feature fusion is proposed.Based on neural network model VGG19,a top-down feature path and a bottom-up attention path are constructed to preserve low-level features while carrying out convolution,which is more conducive to the task of smoke recognition.In this study,the smoke image data set published by the State Key Laboratory of Fire Science was used for experiments,and the accuracy rate reached 99.75%,which was 9.85%higher than that of only using the neural network model VGG19.
作者
周诗怡
李桂梅
黄佳辉
刘坤
王莉
阙帆辉
谢思源
ZHOU Shi-yi;LI Gui-mei;HUANG Jia-hui;LIU Kun;WANG Li;QUE Fan-hui;XIE Si-yuan(School of Intelligent Engineering and Intelligent Manufacturing,Hunan University of Technology and Business,Changsha 410205,China;College of General Education,Hunan University of Information Technology,Changsha 410151,China)
出处
《湖南师范大学自然科学学报》
CAS
北大核心
2024年第5期127-133,142,共8页
Journal of Natural Science of Hunan Normal University
基金
湖南省自然科学基金资助项目(2023JJ50017)
湖南省教育厅科学研究项目重点课题(22A0464)。
关键词
图像处理
火灾烟雾识别
特征金字塔
注意力机制
image processing
fire smoke identification
feature pyramid
attention mechanism