基于深度学习的两级室内火灾检测研究  

Research on Two-level Detection of Indoor Fire Based on Deep Learning

作　　者：王晨旸 WANG Chenyang(Department of science and Technology,Jianghuai College,Anhui University,Hefei 230031,China;Expetimental and practical training center,Hefei University of Economics,Hefei 230031,China)

机构地区：[1]安徽大学江淮学院理工部,安徽合肥230031 [2]合肥经济学院实验与实训中心,安徽合肥230031

出　　处：《宿州学院学报》2023年第3期20-23,共4页Journal of Suzhou University

基　　金：安徽大学江淮学院科研项目(2020KJ0001)。

摘　　要：防止室内环境中火灾的发生与蔓延,降低火灾给人们生命财产安全带来的损失,对室内进行火焰检测是十分必要的。鉴于传统火灾识别过程中需要人工提取特征,且检测过程易受环境干扰,因而提出一种基于深度学习的火灾识别算法。首先将深度学习与计算机视觉相结合,提出一种基于两级分类器的火灾图像判断检测方法。为提高模型训练的泛化能力,在已有数据集的基础上使用GAN生成了大量数据,利用HOG+Adaboost分类器具备高召回率的特性对可能存在的火灾情况进行初判,利用CNN+SVM高精确度分类器对火灾区域进一步次级识别提高识别精度。实验结果表明,火灾图像识别方法较其他算法而言能够以较少的样本,经训练后获得较高的识别准确率,同时该方法对样本训练及检测所需硬件环境要求不高,训练环境也有明显优势。该方法对火灾图像的识别率可达92%以上,识别样本图片平均时间仅需0.87 s,具有较高的有效性及鲁棒性。In order to prevent the occurrence and spread of fire in the indoor environment and reduce the loss of people′s lives and property caused by fire,it is necessary to detect indoor flame.Therefore,indoor flame detection is very necessary.In view of the need for manual feature extraction in the traditional fire identification process,the detection process is easy to be disturbed by the background.Therefore,a fire recognition algorithm based on deep learning is proposed.This paper combines deep learning with computer vision to propose a fire image judgment method based on two-level classifier.In order to improve the generalization ability of model training,a large amount of data is generated by using GAN on the basis of existing data sets.Then,the feature of HOG+Adaboost classifier with high recall rate is used to prejudge the possible fire situation and the high accuracy classifier CNN+SVM is used for further secondary identification in the fire area to improve the identification accuracy.The experimental results show that compared with other algorithms,the proposed fire image recognition method can obtain higher recognition accuracy after training with fewer samples.Meanwhile,the proposed method does not require high hardware environment for sample training and detection,and the training environment also has a distinct advantage.The recognition rate of fire images can reach more than 92%by using this method,and the average recognition time of sample images is only 0.87 s.It has high validity and robustness.

关 键 词：机器学习 GAN CNN SVM 火灾识别 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TP391.41[自动化与计算机技术—控制科学与工程]