基于迁移学习的火灾图像检测方法研究  

Research on Fire Image Detection Method Based on Transfer Learning

作　　者：陈照悦 张红梅[1] 张向利[1] CHEN Zhao-yue;ZHANG Hong-mei;ZHANG Xiang-li(School of Information and Communication Engineering,Guilin University of Electronic Technology,Guilin 541004)

机构地区：[1]桂林电子科技大学信息与通信学院,桂林541004

出　　处：《现代计算机》2020年第28期23-28,共6页Modern Computer

基　　金：国家自然科学基金(No.61461010)。

摘　　要：当前基于特征的火灾检测方法存在误报率高、实时性差等问题,而基于深度学习的火灾检测方法也存在数据集少、预测速度慢、准确率低等问题。针对这些问题,提出一种基于迁移学习的火灾图像检测方法,将源域中训练好的模型,迁移到火灾检测领域。首先,从网络获取InceptionV3、ResNet18、ResNet50、DenseNet121这四种预训练模型,然后将预处理好的训练集放到预训练模型中进行训练,最后对模型进行微调。实验结果表明,DenseNet121相比于其他三个模型具有更好的识别能力,其中准确率达到92.54%,漏报率达到7.36%,且预测时间减少40%,模型大小只有31.6M。The current feature-based fire detection methods have the problems of high false alarm rate and poor real-time performance,while the deep learning-based fire detection methods also have the problems of few data sets,slow prediction speed,and low accuracy.In response to these problems,a fire image detection method based on transfer learning is proposed,which transfers the trained model in the source domain to the field of fire detection.First,get the four pre-trained models InceptionV3,ResNet18,ResNet50,DenseNet121 from the network,then put the pre-processed training set into the pre-trained model for training,and finally fine-tune the model.The experimental results show that DenseNet121 has better recognition ability than the other three models.Among them,the accuracy rate is 92.54%,the false negative rate is 7.36%,and the prediction time is reduced by 40%.The model size is only 31.6M.

关 键 词：火灾检测 迁移学习 微调 DenseNet 

分 类 号：X932[环境科学与工程—安全科学] TP391.41[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术]