Research On Building Seismic Damage Detection Method Based On Deep Learning

Earthquake disaster is considered to be one of the most hazardous natural disasters in human survival.Architecture is an important characteristic of the survival of disaster destruction and the survival of mankind,which has an important real sense for rescue and rebuilding after disaster.During the extraction of architectural disaster information,the traditional artificial design method extracts a single seismic characteristics and has a low accuracy,but the seismic detection method based on the depth learning theory is able to learn more high-quality seismic characteristic expression from unprocessed epicenter image data,and provided a new technical concept for earthquake-related images.For this reason,the text uses a depth learning method to complete the detection of earthquake damage to the building,and enhances the accuracy and automation of the building earthquake injury detection.The main work in the text is as follows:(1)In order to achieve quick detection of the characteristics of earthquake damage,two datasets of drone remote sensing images were binding based on a regression detection model,and a overall detection plan for architectural earthquake damage was designed.A visualization dictionary model consisting of texture,geometric characteristics,key and collapse rate was constructed,and seismic damage was classified in a hierarchical manner.Considering the diversity of earthquake injury characteristics in response to the "overfit" problem due to lack of earthquake injury datasets,designed a dataset expansion algorithm that combines traditional data enhancement and CGAN models to provide data basis for subsequent earthquake damage detection.(2)As for the problems of low correlation degree caused by local convolution of network model and insufficient extraction of seismic damage features caused by dense buildings in remote sensing images,Swin Transformer module is introduced to enhance the connection between images.By basing on the geometric shape,collapse rate and texture information in architectural disasters,important multiple attention mechanism in the Swin Transformer module.A value matrix is constructed,and it is introduced to the neck of the regression model to achieve accurate extraction of the earthquake.Designed a 4-layer BIFPN network structure to solve the problem of errors testing due to the confusion between the high-density building and collapsed building and the background,enriching the shallow layer of meaning information in the image,and replace the CAMB module in the neck of regression detection model,the horizontal scale connection and the two-way passage were combined to improve the accuracy of the earthquake.(3)In order that satisfy the needs of real-time detection of uncollapsed buildings,a lightweight complex module was designed on the feature extraction part of the detected network principal.Shufflenet V2+Stemblock reduced the number of network parameters and increased the execution speed to detect the network model.In response to the lack of extraction of earthquake information due to the diversification of earthquake-related information in buildings,we propose a way to fusion of a passage,space area,and a multi-scale characteristic fusion in the CAMF module.The method was introduced to the end of the detection network backbone to efficiently extract the characteristics of earthquake-related characteristics.Finally,the combination of loss functions based on Logsoft Max+NLLloss is designed to solve the problem of detection accuracy due to the slope loss of traditional loss functions and the intersection of architectural classifications that had not collapsed.Finally,for the sake of proving the effectiveness of the proposed method,we conducted an experiment on earthquake disaster detection of buildings that have the crude earthquake damage detection and not collapsed on the Pytorch platform,respectively.As a result,this method has been revealed that for the earthquake damage detection of buildings.