Research On Bridge Crack Identification And Measurement Method Based On Convolutional Neural Network

The total number of Chinese bridges ranks first in the world.In recent years,bridge accidents have occurred frequently,which not only affects traffic,but even endangers people’s lives.In order to improve the detection level of bridge cracks and solve the problem of time-consuming and labor-intensive in manual detection and the problem of traditional image processing methods need to be set manually.Here in the deep learning technology is used as the theoretical basis.A method for bridge crack classification,location and measurement based on convolutional neural network is proposed.The main contents of this paper include:1.Aiming at the lack of crack label sample data,a large-scale bridge crack original dataset is constructed.Then the characteristics of the bridge crack image are analyzed,and the images of the original data set are processed by the Multi-Scale Retinex algorithm with color restoration factor,graying,improved bilateral filtering algorithm,and HEL(Histogram Equalization-Laplace)feature enhancement algorithm.A pre-processed bridge crack dataset is constructed and named as RLH(Retinex-Laplace-Histogram Equalization)dataset,which is prepared for the subsequent training.2.Aiming at the problem that traditional image processing algorithms need to manually set parameters and the problem that the classic deep learning model is used to detect bridge cracks with low accuracy,a bridge crack detection algorithm based on improve-Goog Le Net network is proposed.An image classification system for bridge crack characteristics is established.3.The sliding window is used to locate the crack,and the precise positioning of the crack is realized.Combined with the skeleton extraction algorithm to calculate the length and width of the crack,effective measurement of cracks is achieved.The experimental results show that compared with the original data set,the RLH data set processed by the preprocessing algorithm in this paper improves the recognition accuracy by 2%.Compared with the original Goog Le Net network,the improve-Goog Le Net network improve the recognition accuracy by 3.13%,the training time drop to the 64.6% of the original training time.In addition,the trend of cracks is considered in the skeleton extraction algorithm,and the width can be calculated more accurately.Moreover,both the maximum width and average width can be calculated.The average error of length extraction and width extraction is 2.87% and 4.29% respectively.In summary,the classification,positioning and measurement methods proposed in this paper have the advantages of high accuracy and fast speed.