基于改进Resnet18网络的火灾图像识别  

Fire image recognition based on improved Resnet18 network

作　　者：陈跨越 王保云 Chen Kuayue;Wang Baoyun(School of Mathematics,Yunnan Normal University,Kunming 650500,China;Yunnan Key Laboratory of Modern Analytical Mathematics and Applications,Yunnan Normal University,Kunming 650500,China)

机构地区：[1]云南师范大学数学学院,昆明650500 [2]云南师范大学云南省现代分析数学及应用重点实验室,昆明650500

出　　处：《河南师范大学学报（自然科学版）》2024年第4期101-110,I0013,I0014,共12页Journal of Henan Normal University(Natural Science Edition)

基　　金：国家自然科学基金(61966040).

摘　　要：针对传统卷积神经网络进行火灾图像识别时,准确率不高、特征难以提取、网络的平移不变性较弱等问题,对Resnet18网络进行改进,使其具有更高的性能和准确性.首先,在Resnet18网络的卷积层前插入空间变换网络(spatial transform networks,STN).对于连续多个相同大小的卷积层,只在第一个卷积层前添加STN,共添加5个,并且在全连接层后添加dropout层防止过拟合.然后,使用迁移学习(transfer learning,TL)的方法对火灾进行分类识别.实验结果表明,改进后的Resnet18网络准确率、召回率、F_(1)值和AUC值等各项指标性能优于Resnet18网络和其他深度学习识别算法,能够对火灾图像进行快速、准确地识别.In view of the problems such as low accuracy,difficult feature extraction and weak translation invariance of the network during fire image recognition by traditional convolutional neural network,this paper improved Resnet18 network to make it have higher performance and accuracy.First,the spatial transformation network(STN)is inserted in front of the convolution layer of the Resnet18 network.For multiple convolution layers of the same size in a row,only the STN is added before the first convolution layer,a total of five are added,and the dropout layer is added after the fully connected layer to prevent overfitting.Then,the transfer learning(TL)method is used to classify and identify fires.Experimental results show that the improved Resnet18 network accuracy rate,recall rate,F_(1)value and AUC value are superior to Resnet18 network and other deep learning recognition algorithms,and can quickly and accurately identify fire images.

关 键 词：火灾检测 卷积神经网络 空间变换网络 Resnet18 HSI色彩模型 迁移学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术] X928.7[环境科学与工程—安全科学]