Research On Crack Detection Algorithm Of Concrete Structure Based On Computer Vision

During the structure damage detection after an earthquake event,the crack morphology of concrete structures not only represents seismic action but also reflects the position of the internal force and the direction of the internal stress.It can be seen that correctly measuring the width,length,direction and other parameters of concrete structure crack plays an important role in scientifically evaluating the bearing capacity state of the structure.At present,estimate the damage level of structures at the earthquake site mainly rely on the experts with experience which is time-consuming,laborious and rough.With the development of deep learning and image processing technology,it will be the trend of quick evaluation of structural damage in the future to combine the two and evaluate the structural damage by using the images taken in the earthquake site.In order to make full use of computer vision technology to solve the problem in damage detection of concrete structures,a set of algorithm for crack detection and quantitative analysis based on deep learning and traditional image processing technology is proposed.The main work and conclusions are as follows:Firstly,the latest damage detection based on computer vision technology,emphatically the computer vision technology based on deep learning in the study of damage detection are summarized.The existing and latest damage data sets are introduced.Each dataset has their standards.According to the use of different learning theory,it can be divided into three categories as followings: 1)based on the classification task,2)based on the object frame,3)based on the three kinds of semantic segmentation.At last,the advantages and disadvantages of every kind of damage detection methods are specifically analysed.Secondly,aiming at the problem that the seismic images can't be directly used to make dataset and to segment the cracks,a series of pretreatment to preprocess the original images was researched.Firstly,transform the oblique shadow image into orthogonal projection image,and then to remove blue handwriting.Then in order to remove blue handwriting by using the continuity of the HSV color space,the original image is projected to HSV color space.Finally,the RGB image is transformed to gray level image and cut into 512 pixels by 512 pixels by sliding window method.This series of pretreatment provided a standard for making a damage detection datasets.Thirdly,aiming at the problem that the crack segmentation algorithm based on traditional image processing is tedious and the generalization ability of the algorithm is weak,an algorithm based on fully convolutional neural networks is researched.Our own dataset is made by using subimages obtained from preprocessing.In order to validate the capacity of our algorithm,the segmentation results with traditional image segmentation algorithms like iterative threshold method,the mathematical morphology operator and shallow neural network is been compared with.The experiment shows that this algorithm has better segmentation accuracy and efficiency.Fourthly,aiming at quantifying the characteristics of concrete cracks,an algorithm is researched to analysis the crack.Three image processing methods named connected domain denoising,fracture edge extraction and fracture skeleton extraction were used to further process the segmented crack area.The above process enabled the crack quantization algorithm to more accurately measure the characteristics of crack direction,width and length.Experimental results show that the fracture feature algorithm can effectively measure the direction,width and length of concrete cracks.