基于经典深度卷积神经网络算法的火灾图像识别方法  

Fire Image Recognition Method Based on Classical Depth Convolution Neural Network Algorithms

作　　者：何豪 王杰军 HE Hao;WANG Jiejun(School of Electrical and Information Engineering,Tianjin University,Tianjin 300072;Siterwell Electronics Co.Ltd.,Ningbo 315031)

机构地区：[1]天津大学电气自动化与信息工程学院,天津300072 [2]赛特威尔电子股份有限公司,浙江宁波315031

出　　处：《常州工学院学报》2023年第4期20-25,共6页Journal of Changzhou Institute of Technology

基　　金：浙江省安全工程与技术研究重点实验室开放基金项目(202103)。

摘　　要：基于4种经典深度卷积神经网络(DCNN)算法模型,在火灾图像识别方面进行了应用实践。建立火灾图像数据集,使用Tensorflow框架搭建训练环境,比较4种算法模型在训练和识别过程中的差异性。结果表明:SqueezeNet算法具有训练时间短、模型文件小等优点,但训练精度较其他算法有所降低;Inception算法综合表现好,具有较高的准确率和中等的训练时间,而且收敛速度快;4种DCNN算法对多数火灾场景的识别准确率较高,但在对非火灾场景的抗干扰性方面存在较大提升空间。Four classical deep convolutional neural network(DCNN)algorithm models based on fire image recognition have been applied in practice.A fire image dataset is built,a training environment is constructed using the Tensorflow framework,and the differences of the algorithmic models in the training and recognition process are compared.The results show that:the SqueezeNet algorithm has the advantages of shorter training time and smaller model files,but the training accuracy is reduced compared with other algorithms;the Inception algorithm performs well overall,with its higher accuracy and moderate training time,and faster convergence;the four DCNN algorithms have a high recognition accuracy for most fire scenarios,but there is more room for improvement in terms of anti-interference performance of non-fire scenes.

关 键 词：深度学习 卷积神经网络 火灾图像 识别准确率 火灾探测方法 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]