Research And Realization Of Seismic Damage Information Recognition Of Buildings Based On Deep Learning

In earthquake disaster,building damage is the main cause of economic loss and casualties.It is very important to identify the information of building earthquake disaster quickly for the reasonable arrangement of disaster relief and reconstruction.In this paper,building damage information recognition of high-resolution remote sensing image after the earthquake is taken as the research purpose,and the deep learning method is combined with the high-resolution remote sensing image after the earthquake to realize the rapid and accurate recognition of building damage information after the earthquake.In this paper,the methods of deepening network structure,improving convolution mode and adding batch standardization layer are integrated to explore the optimization method of Unet network model and its application in building seismic damage information identification.Taking the Kumamoto earthquake image as the sample data,it is proved by experiments that the improved Unet structure has achieved good results in the post earthquake building segmentation,and the accuracy and speed of the segmentation results are improved compared with the Unet network under the same conditions,and the MIo U of the segmentation results can reach 0.8109.Finally,this paper constructs a high-resolution remote sensing image building seismic damage information extraction system based on Browser / server mode.The system sub module realizes the functions of sample making,model training,earthquake damage identification,and simplifies the operation process of using depth learning to segment post-earthquake buildings.The research content includes the following four aspects:(1)First of all,this paper analyzes the research status of building seismic damage extraction methods based on remote sensing image,and focusing on the extraction methods based on multi temporal and single temporal remote sensing images,as well as the development of deep learning methods in the field of building seismic damage extraction.(2)In view of the lack of remote sensing image sample set of building seismic damage information,based on Kumamoto earthquake image,from the aspects of image preprocessing,sample annotation,data augmentation and so on,this paper puts forward the flow operation of making semantic segmentation data set.(3)Because the structure of the Unet model is clear,easy to train,and the excellent performance of the structure in the field of building extraction,based on the Unet segmentation model,this paper explores an improved Unet network model which is more suitable for identifying building seismic damage.The network structure depth,convolution mode and batch standardization layer are optimized to improve the accuracy and operation speed of building seismic damage identification.(4)Based on the improved Unet structure,a building seismic damage extraction system is constructed.The system integrates image preprocessing,sample set making,model training and earthquake damage identification module.It is a flow operation system from image input to output of extraction result image.