基于UO-Net模型的火焰检测方法  

Flame Detection Method Based on UO-Net Model

作　　者：陈浩霖 高尚兵 相林[1] 严云洋 黄子赫 蔡创新[1] CHEN Haolin;GAO Shangbing;XIANG Lin;YAN Yunyang;HUANG Zihe;CAI Chuangxin(School of Computer and Software Engineering,Huaiyin Institute of Technology,Huai’an 223001,China;Laboratory for Internet of Things and Mobile Internet Technology of Jiangsu Province,Huaiyin Institute of Technology,Huai’an 223001,China;School of Computer Engineering,Jiangsu Ocean University,Lianyungang 222005,China)

机构地区：[1]淮阴工学院计算机与软件工程学院,江苏淮安223001 [2]淮阴工学院江苏省物联网移动互联技术工程实验室,江苏淮安223001 [3]江苏海洋大学计算机工程学院,江苏连云港222005

出　　处：《江苏海洋大学学报（自然科学版）》2020年第4期8-15,共8页Journal of Jiangsu Ocean University:Natural Science Edition

基　　金：国家重点研发计划项目(2018YFB1004904);江苏高校“青蓝工程”项目;江苏省高校自然科学研究重大项目(18KJA520001);江苏省“333工程”资助项目(BRA2016454);淮安市科技局计划项目(HAB201803)。

摘　　要：一般基于深度学习的火焰检测方法识别效率不够理想,主要原因是特征提取网络中参数较多。对此,提出了一种基于UO-Net模型的火焰检测方法。UO-Net模型是在YOLOv3模型基础上建立的一种端到端的深度神经网络模型。该方法采用多卷积核组合结构,减少特征提取网络层的通道数。同时,提出了一种图像分割网络来加快模型的训练速度,并利用图像分割网络的注意力图来指导检测模型对火焰进行检测,从而提高火焰检测模型的性能。在Bilkent大学公开的火焰数据库VisiFire和真实场景数据集上对UO-Net模型进行了测试,最终识别准确率和帧率分别达到96.6%和42 f/s。实验结果表明,该方法能够从轻量级网络中提取火焰特征,检测速度和检测精度均优于现有的其他方法。Generally speaking,the detection efficiency of flame detection methods based on deep learning is not ideal,because there are many parameters in the feature extraction network.To solve these problems,the flame detection method based on UO-Net model was proposed in this paper.Method The UO-Net model is an end-to-end deep neural network model based on the YOLOv3 model.This method used a combination of multi-kernel convolution to reduce the number of channels in the feature extraction network layer.At the same time,the image segmentation network was proposed to speed up the training of the model,and in order to improve the performance of the flame detection model,the attention map of image segmentation network was used to guide the flame detection.Result UO-Net model was tested on the visifire sets of the fire database of Bilkent University and real scene data sets.The final detection accuracy reached 96.6%with a frame rate of 42 f/s.Experimental results show that this method can extract flame features from a lightweight network,and the detection speed and detection accuracy are better than existing methods.

关 键 词：YOLOv3模型 深度神经网络 多卷积核组合结构 特征提取 图像分割 注意力图 火焰检测 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]