Research On Extraction And Assessment Method Of Building Earthquake Damage Based On Semantic Segmentation

With a vast territory,China is one of the countries suffering the most natural disasters in the world.Earthquake is a long-term frequent natural disaster,and its instantaneous suddenness and strong destructive power pose a serious threat to the safety of people’s lives and urban buildings.Therefore,carrying out the postearthquake damage assessment of urban buildings is of great significance for postdisaster emergency rescue,social order restoration and people’s livelihood security.How to achieve accurate,comprehensive and rapid earthquake damage extraction and damage level assessment of urban buildings after an earthquake is a long-term challenge in the field of engineering structure disaster prevention and mitigation.At present,building earthquake damage assessment mainly relies on professional field visual survey and seismic response monitoring data to calculate damage index and other methods.At present,the rapid development of deep learning theory and computer vision technology has provided important support for mining the depth characteristic information of earthquake damage of building structures and realizing the precise and quantitative extraction and evaluation of damage.Aiming at the characteristics of various types of earthquake damage of building structures,underbalanced feature information,geometric multi-scale,multi-factor interference,and complex scenes,this paper proposes a refined quantitative extraction and evaluation classification method for seismic damage inside building structures based on the principle of pixel-level semantic segmentation.The research contents are as follows:（1）A semantic segmentation and extraction method for the damage of hierarchical cracks in structural components of complex scenes is proposed.Using a dataset of 3543 representative crack damage images of structural components at home and abroad,the hyperparameter training and improvement and optimization of the UNet semantic segmentation network model were carried out,and the simple,difficult to distinguish by human eyes,complex lines,strong interference backgrounds,etc.were realized.The classification cracks have good identification accuracy and prediction effect.Among them,the overall average m Io U value of the validation set data can reach 85.54%,the average m Io U value of the test set is 83.69%,and the highest m Io U value in all scenarios can reach 90.1%.（2）A data augmentation and refined extraction method for unbalanced information of four types of coupled failure features,namely "spalling","dropping block","rebar exposure" and "rebar buckling",was developed.Selected 13354 image data sets of building component damage covering Wenchuan and other domestic representative earthquakes,established a semantic segmentation model of multicategory coupled earthquake damage based on the improved U-Net network,and adjusted the network model loss function weight coefficient and introduced The "mosaic" method enhances the unbalanced characteristic information data of steel bars,and realizes the high-precision extraction and identification of multiple types of seismic coupling damage.Among them,the predicted failure Io U values of the“peeling” and “dropping” categories with rich feature information are 91.95% and92.11%,respectively;while the two categories of “rebar exposure” and “rebar buckling” have less damage categories,respectively from the original 72.53% and73.55% increased to 89.71% and 88.53%.（3）A preliminary method is established from the structural component level seismic damage to the structural seismic damage level assessment.Combined with the existing post-earthquake structural damage assessment criteria and component destructive experimental image data calibration,by extracting five types of failure geometric parameters（such as crack shape,spalling area ratio,etc.）The semiquantitative mapping relationship between the damage damage index and the image damage category is used to preliminarily explore the evaluation method of earthquake damage level from component damage to structural monomer.