台风灾后存留房屋安全隐患自动辨识及位置定位研究  

Research on automatic identification and localization of hidden hazards in houses remaining after typhoon disaster

作　　者：段在鹏 黄豪琪 孙文磊 朱俊杰 DUAN Zaipeng;HUANG Haoqi;SUN Wenlei;ZHU Junjie(School of Economics and Management,Fuzhou University,Fuzhou 350108,China;Fujian Emergency Management Research Center,Fuzhou 350108,China;College of Environment and Safety Engineering,Fuzhou University,Fuzhou 350108,China;School of Computer and Information Engineering,Chuzhou University,Chuzhou 239000,Anhui,China)

机构地区：[1]福州大学经济与管理学院,福州350108 [2]福建省应急管理研究中心,福州350108 [3]福州大学环境与安全工程学院,福州350108 [4]滁州学院计算机与信息工程学院,安徽滁州239000

出　　处：《安全与环境学报》2024年第11期4232-4243,共12页Journal of Safety and Environment

基　　金：国家社会科学基金项目(23BGL290)。

摘　　要：台风及其引发的洪涝灾害会对大量房屋造成损伤,灾后存留房屋排查任务又急又重,而人工排查房屋隐患耗时费力,研究提出一套可对台风灾后存留房屋隐患部位自动辨识、定位,并对房屋安全水平自动分级的模型。首先,采集台风“海葵”后福州受灾房屋图像,以数据增强的方法使3个安全等级房屋样本达到基本均衡的水平;其次,通过试验对比得到更适合作为ConvNeXt网络对房屋安全分级的图像尺寸处理方式,并进一步通过学习率预热机制提升模型分类准确率;再次,以消融试验探究门控通道注意力(Gated Channel Transformation,GCT)嵌入ConvNeXt的相对最佳的方式;最后,利用Ablation-CAM(Ablation Class Activation Mapping)技术提供模型对隐患位置定位的可视化结果,验证模型辨识台风灾后房屋隐患的准确性。试验结果显示,研究提出的模型在台风灾后存留房屋安全分级任务中准确率达到了93.87%、F_(1)为93.95%,均高于其他主流网络;同时,与ConvNeXt相比,GCT-ConvNeXt的隐患位置定位结果更加精准,可为台风灾害后房屋隐患排查工作提供参考。The typhoons,accompanied by the floods they unleash,have the potential to inflict extensive damage on a significant number of residential structures.The inspection task of post-disaster remaining houses is both urgent and heavy,while manually inspecting hidden dangers in houses is time-consuming and laborious.This paper proposed a model that can perform automatic hazard identification,localization,and safety grading of post-typhoon remaining houses.First,images of houses affected by Typhoon Haiyan in Fuzhou were collected,and the data augmentation technique was employed to achieve a basic balance in the samples of houses with three safety levels.Second,through comparative experiments,an optimal image size processing approach suitable for ConvNeXt-based house safety grading was determined.Specifically,directly resizing both the long and short sides of the image to 224 pixels can better improve the classification accuracy of ConvNeXt.Building on this,the model's recognition accuracy was further enhanced by the utilization of the learning rate warmup mechanism.Thirdly,the relatively optimal way of embedding Gated Channel Transformation(GCT)into ConvNeXt was explored by ablation experiment.It is found that embedding the GCT after the depthwise convolution layer in each ConvNeXt block results in the best performance for GCT-ConvNeXt,without significantly increasing the training time of the model.Finally,the Ablation Class Activation Mapping(Ablation-CAM)technique was applied to provide visualization results of the model's localization of hidden hazards,verifying the accuracy of our model in identifying hazards in post-typhoon remaining houses.The experimental results show that the proposed GCT-ConvNeXt reaches 93.87%and 93.95%in accuracy and F 1 score respectively in the post-typhoon remaining house safety grading task,outperforming other mainstream networks.Additionally,GCT-ConvNeXt demonstrates high efficiency in identifying the safety level of houses,and its hazard localization results are even more accurate,servin

关 键 词：安全工程 台风灾害 房屋安全 隐患排查 深度学习 可视化技术 

分 类 号：X947[环境科学与工程—安全科学]