Research On Concerete Crack Detection Algorithm Based On Deep Learning And 3D Reconstruction Techiniques

Crack is one of the most common damages of concrete structures.Timely and accurately detect crack is significant to assess the safety and reliability of structures.With the construction and long-term service of a large amount of civil insfrastructures in the past decades,concerete crack detection based on manully inspection is incompetence for settling such great demands growing.Manully inspection have the following drawbacks such that the decision is subjective according to the inspector’s expertise,high labor intensity,and low working efficiency leading the research of automatic and intelligent crack detection algorithm is in full swing.However,attribute to the change of measurement environment,sensor error,low quality of detection data and detection algorithm error,the existing crack detection methods are ineffective to meet the precision requirements of practical engineering.Therefore,according to structural surface image information and advanced computer vision technology,this paper developes an automatic and intelligent crack detection algorithm based on deep learning and 3D reconstruction technology,to improve the effectiveness and robustness of intelligent crack detection method and promote the application of automatic detection technology in engineering practice.The main work and achievements are as follows:1.Through the study of deep learning algorithms in computer vision,this paper proposed28 network models for concrete crack recognition,combing convolutional neural network,transformer and object space detail compensation technology such as skip connection,feature fusion and feature pyramid netowrk.After the design of network module experience,three concrete crack detection network models FPAFFN1,2 and 3 with different encoder configurations were screened out from the 28 network models.Through the comparison experiment with open dataset and comparison experiment of the model parameters in different training phases,crack identification performance of above network models are studied,and the optimal configuration is determined as FPAFFN3.FPAFFN3 achieved 75.25% mean Intersection over Union(m Io U)index value in CRACK500 data test set,which was at most10% and at least 2% more than other downstream networks.The m Io U values of FPAFFN3 for test results are also the best in other open crack data sets,with demonstrate remarkable robustness.2.Based on the lightweight optimization methods,there was a network model parameter change analysis for FPAFFN2 and 3.Considering the number of parameters,floating point of operations(FLOPs)and Graphics card memory usage,three lightweight optimization schemes are proposed to guide the backbone model optimization.Three experience: model baseline optimization,model depth optimization,model channel optimization were respectively carried out according to the optimization schemes.Through the results verify of the model baseline optimization experience,based on FPAFFN2 and 3,FPAFFN4 were obtained.FPAFFN4 uses depthwise separable convolution to reduce the number of parameters and FLOPs by more than 90%.The identification accuracy of FPAFFN4 is 0.7% m Io U worse than FPAFFN3 with the compromise between the number of parameters and the identification accuracy.Through model channel optimization experience,FPAFFN5 developed based on FPAFFN2 reduces 90% of the original parameter number,80% of FLOPs and half of memory usage.The performance of FPAFFN5 is slightly worth than FPAFFN3,which is the best lightweight crack recognition network in comprehensive consideration.3.To effectively identify cracks and accurately locate the cracks in the three-dimensional space of structural components to further reduce manual participation and improve the efficiency of structural health detection,a three-dimensional crack damage identification system is constructed by combining FPAFFN5 and three-dimensional reconstruction algorithm.In the experiment of crack detection with an engineering concrete slab,the pose guidance of the data collected by the camera is obtained.The system is capable of accurately identify component cracks from the pictures and build a three-dimensional damage model,including the complete crack information.The experimental results verify the effectiveness and robustness of FPAFFN5 and the three-dimensional crack damage identification system in practical engineering.