Research On Concrete Crack Detection And Evaluation Based On Deep Learning

Concrete crack diseases are not only the main form of building surface damage,but also the focus of building damage assessment and maintenance.With the increasing use time of earlier buildings,the number and distribution area of building diseases are increasing,and building health monitoring and maintenance has gradually become a research hotspot.Cracks are one of the most common,easily occurring and harmful diseases in buildings.Long service time,processing defects and weather changes may lead to small cracks in buildings.The threat of cracks to safety is not the most serious.However,if not treated in time,cracks may lead to other more serious secondary diseases.Once the cracks are built,they will cause more serious secondary diseases The collapse of buildings will cause serious safety accidents.Therefore,it is necessary to detect and repair the building cracks in time,so as to prevent the damage from further expanding.The traditional manual maintenance method has been widely used in the past,but its detection efficiency is low with strong subjectivity,which is not suitable for the increasing maintenance workload.At present,the existing building crack detection algorithms generally have many problems,such as low technical universality and low detection efficiency.A detection algorithm can only be used for specific situations,It is difficult to adapt to all kinds of complex and changeable building surface background,so it needs further optimization and improvement.In order to solve the above problems,based on the analysis of the existing crack detection literature,this paper proposes a crack detection algorithm based on deep learning(1)Aiming at the problem of complex background and changeable texture of cracks in natural environment,this paper proposes a hole convolution pyramid densenet sub block crack detection and classification algorithm to initially locate the crack area.Firstly,the crack feature extraction network densenet is proposed.Densenet uses dense connection to extract convolution features,and makes full use of the features of each sub layer in the convolution layer,which can improve the quality of cracks Then,a squeeze excitation module is proposed to learn the features extracted from the network.The importance of each channel in the original feature map is evaluated.According to the importance,the channel features are weighted out to obtain more effective crack features.Finally,a hole convolution pyramid module is proposed to learn by fusing the context information,and the holes with different sizes are extracted The hole convolution has different size receptive field,and classifies the sub blocks from the multi-scale perspective,which can effectively reduce the misclassification of complex background and improve the accuracy of sub block detection.(2)Aiming at the problem that the traditional crack image segmentation methods can not accurately extract cracks,this paper proposes a whole nested network concrete crack segmentation method based on convolution deconvolution feature fusion.Based on the research of coding decoding architecture,the proposed network first uses vgg-16 as the basic feature extraction network,aiming at the problem of feature redundancy in coding stage,based on channel attention mechanism,uses channel spatial correlation and global information to stimulate crack features and remove redundant features;secondly,deconvolutes deep semantic information through convolution deconvolution feature fusion module Secondly,based on the multi-scale supervised learning mechanism,the holistically nested networks is used to solve the problem The prediction results of different scales are fused to enhance the expression ability of the network to the linear topological structure and improve the accuracy of fracture segmentation.Finally,a hybrid void convolution boundary refinement module is proposed to further refine the fracture boundary and improve the accuracy of fracture segmentation.At bridge＿Crack＿Image＿A large number of experiments on data sets and CFD data sets show that,compared with other depth networks,the proposed network not only has better segmentation results for cracks with different widths,but also has stronger robustness.(3)Aiming at the problem that crack detection in real scene needs to take on-the-spot photos and calculate crack parameters,this paper proposes an evaluation index of building surface damage.Firstly,a crack thinning algorithm based on look-up table method is adopted to obtain the complete skeleton of crack,and then a set of calculation method of crack geometric parameters is developed based on image processing theory,including crack area and crack length Finally,an online crack detection app based on B / S architecture is designed to realize the detection algorithm proposed in this paper,which is convenient for online crack sampling,automatic calculation of crack parameters and query of detection history...