Damage Assessment Of Buildings Based On Multi-view Stereo Deep Learning Network After Earthquake

After the earthquake disaster,timely and accurate assessment of damaged buildings is crucial.Because it is directly related to the safety of people ’s lives and property.At present,the method based on manual field investigation is often time-consuming and laborious.Most of the existing researches based on artificial intelligence(AI)methods use two-dimensional information of orthophotos to evaluate damage,but this method cannot provide multi-view information of damaged buildings,resulting in inaccurate classification results.Therefore,this paper proposes a MVS(multi-view stereo)model based on deep learning to assist the classification of post-earthquake damaged buildings.The method includes three main processes.Firstly,the camera parameters of each image need to be calculated;then,the accurate and efficient Cas MVSNet algorithm is used for3 D reconstruction.Finally,based on the 3D model,the building damage assessment classification is performed.In this paper,the method is tested on the UAV aerial image acquired by Yangbi,Yunnan Province to evaluate its effectiveness.The 3D model of damaged buildings after the earthquake show that: Firstly,compared with other deep learning models and commercial software,Cas MVSNet has obvious time efficiency advantages and can meet the timeliness requirements of post-earthquake rescue and loss assessment.In addition,Cas MVSNet has the lowest memory consumption.Secondly,Cas MVSNet shows the best 3D reconstruction results in high-rise and small buildings.Finally,our method can provide detailed information and multi-angle damage information of damaged buildings,which can better assist the building damage assessment task.Compared with the real results of field investigation,our classification results are very close.Therefore,our research results provide a new method for the assessment of damaged buildings,which has high timeliness and accuracy and can provide effective support for rescue operations.The multi-view stereo network based on deep learning has become an important technology in the field of 3D reconstruction.However,in order for this technology to play a greater role in practical engineering applications,many factors need to be considered,including time efficiency,the quality of models,and memory consumption.Therefore,this paper proposes a new network model,which aims to take into account these factors and has a small memory consumption to achieve 3D reconstruction of highresolution images.Specifically,our network model has the characteristics of fast,efficient,accurate and low memory consumption,which can meet the needs of largescale and high-resolution 3D reconstruction of UAV images in post-earthquake areas.Our research focuses on designing a method to optimize the network structure to improve time efficiency and modeling quality,and to improve the performance of the algorithm while maintaining low memory consumption.In the evaluation of test on the public data set DTU,the model built by our network performs best on the integrity index,which proves the effectiveness and reliability of the network.In the self-collected postearthquake data of Jiuzhaigou and post-earthquake data of Yangbi,modeling can effectively reflect the detailed damage information of buildings.Compared with other methods,this paper shows better classification results and accuracy in the classification and evaluation of building damage based on post-earthquake data of Yangbi,which is closer to the real situation of field investigation.In addition to the field of 3D reconstruction,our method can also be applied to other fields,such as urban planning and architectural design.For example,the network can be used to perform threedimensional reconstruction of urban streetscape images to improve the accuracy and efficiency of urban planning.In addition,the network can also be used in architectural design to help architects better predict and evaluate the appearance and structure of buildings.In summary,our research results provide a new method for the optimization of deep learning multi-view stereo networks,aiming to improve their performance in terms of time efficiency,modeling quality,and memory consumption.This method will help to apply this to a wider range of fields and bring greater value to practical engineering applications.