基于目标检测卷积神经网络的图像型火灾探测算法  

Image fire detection algorithms based on object detection convolutional neural networks

作　　者：张苗 李璞 杨漪[3] 宋文华[4] ZHANG Miao;LI Pu;YANG Yi;SONG Wen-hua(Tianjin Heping Fire and Rescue Division,Tianjin 300090,China;Zhengzhou Airport Economy Zone Fire and Rescue Division,He'nan Zhengzhou 450000,China;Xi'an University of Science and Technology,Shaanxi Xi'an 710054,China;School of Environmental Science and Engineering,Tianjin Polytechnic University,Tianjin 300387,China)

机构地区：[1]天津市和平区消防救援支队,天津300090 [2]郑州航空港经济综合实验区消防救援支队,河南郑州450000 [3]西安科技大学,陕西西安710054 [4]天津工业大学环境科学与工程学院,天津300387

出　　处：《消防科学与技术》2022年第6期807-811,共5页Fire Science and Technology

基　　金：天津市科技重大专项与工程(16ZXHLSF00290)。

摘　　要：针对传统图像型火灾探测算法误差率高、延迟探测、计算量大等问题,提出了基于目标检测卷积神经网络(FasterRCNN、R-FCN、SSD和YOLO v3)的图像型火灾探测算法。通过对比实验表明,基于目标检测卷积神经网络的探测算法准确性较高。其中,YOLO v3探测算法的平均精度为84.5%,探测速度为28帧/s,具有更高的稳定性,更适用于图像型火灾探测系统的开发。The existing image fire detection algorithms have the problems of weak generalization ability, high false alarm rate,and low practicality. Based on four advanced object detection convolutional neural networks(e.g. Faster-RCNN, R-FCN, SSD and YOLO v3), new image fire detection algorithms were developed. The comparison of the proposed and current algorithms reveals that the algorithms based on object detection CNNs have significant advantages. Especially, the average precision of the algorithm based on YOLO v3 reaches to 84.5%, and the detection velocity is 28 frame/s. Besides, the YOLO v3 also has stronger robustness of detection performance, and is suitable for developing fire detection system.

关 键 词：卷积神经网络 深度学习 火灾探测 

分 类 号：X932[环境科学与工程—安全科学] X924.4