Research On Damage Detection And Location Method Of Concrete Components Based On Deep Learning

The reasons for the collapse of concrete structures are often due to the lack of scientific and timely disease detection,the cracking of these small structures often leads to catastrophic damage to the whole project.In practical engineering,the traditional detection methods based on visual inspection are usually lack of accuracy and versatility,and can not work under complex infrastructure conditions.Therefore,it is necessary to choose scientific and reasonable methods to check the building structure and regularly evaluate its health status.In addition,an important subsequent task of accurately identifying defects is to find the location of defects and perform specific maintenance.The development of deep learning methods leads to new research on damage detection and image localization technology.This structural damage detection method based on visual elements helps to establish appropriate maintenance strategies,thus saving maintenance budget and improving the performance of infrastructure.Therefore,in order to make full use of deep learning method to solve the detection and localization task of concrete component damage,this thesis designs a method of automatic detection and localization of concrete damage based on computer vision.The main research contents are as follows:In view of the lack of detection standards for various methods caused by the absence of a unified concrete damage dataset,a large-scale concrete damage dataset containing various complex scenarios is constructed.The dataset consists of 10691 images containing typical types of defects such as concrete cracks,spots,rebar exposure and spalling.Aiming at the problems of slow inference speed and low accuracy of concrete defect detection algorithm at present,a novel deep learning method for optimizing concrete damage detection algorithm is studied by using depthwise separable convolution,inverted residual network structure and linear bottleneck structure.Compared with the original concrete damage detection algorithm,the proposed method only accounts for 1/5 of the original network,the network inference speed is increased by 24.1% and 53.5% respectively,and the comprehensive detection accuracy is improved by 3.25%and 4.04%respectively.In order to further improve the performance of concrete damage detection algorithm,concrete damage detection and segmentation algorithms are optimized respectively,and a new two-stage concrete damage detection and segmentation method based on deep learning is studied.The proposed method combines concrete damage detection and segmentation to segment the detected concrete cracks in the second stage,and the detection accuracy is specified to the damage pixels.The experimental results show that compared with the original network,the accuracy of the network is improved by 4.31% and 7.47%,respectively,and the inference speed of the network is improved by 35.3% and 50.3%,respectively.To reduce the cost of data annotation in practical engineering application of concrete defect detection method,improve the accuracy of pixel-level damage detection and realize the location of concrete damage location,a concrete damage detection and location method based on self-supervised instance segmentation and location identification was studied.The model not only has the ability to detect and classify various typical concrete defects in complex scenes,but also can accurately locate the detected defects without external sensors.The proposed model can achieve pixel-level defect detection and geographic location only through the cost of bounding box annotation and automatic indexing.The experimental results show that the detection accuracy and positioning accuracy reach 48.75% and 83.69% respectively,which has significant advantages compared with other methods.