Research On Road Crack Detection Based On Deep Learning

With the continuous development of infrastructure construction in our country today,highway construction,as one of the important links in infrastructure construction,has also been developed rapidly,but the road maintenance problems that come with it.The first step of road maintenance is to detect pavement damage.Among them,road cracks are one of the most common types of pavement diseases in the process of pavement inspection.The thesis takes road cracks as the research content,and uses deep learning semantic segmentation algorithm to detect road cracks.It focuses on the preprocessing of road crack images,crack detection algorithms and quantitative analysis of cracks.First,filter the collected crack images,and combine the characteristics of the crack images,using an algorithm that combines Gaussian bilateral filtering and minimum filtering to remove the interference of noise in the image.Then annotate the image,and use image flip,image translation,and color adjustment to expand the data set for the problem of insufficient sample data.Secondly,the paper compares the seven commonly used algorithm models in the field of neural network semantic segmentation,improves the classic U-Net network model,introduces hole convolution in the encoding and decoding part of the U-Net model,and at each stage of the decoding part.The output is divided into two branches,one is used as the input of the feature map of the next stage,and the other is used as the output of the direct feature map.Then the feature maps of each stage are channel-fused,and finally the result map of the crack segmentation is output.The improved network model optimizes the model structure,reduces the number of parameters of the network model,and increases the detection time of the algorithm.Experimental results show that the improved U-Net network model is superior to other network models in terms of model parameters.The pixel accuracy of the crack reaches98.9%,F1 reaches 86.06%,m Io U also reaches 80.36%,and the model parameters are 13.1M,compared with U-Net model,reduced model parameters by 58%.When applied to the other two public data sets,it also obtained higher experimental results than other methods.Finally,the thesis carried out a quantitative analysis of the cracks,extracted the crack skeleton using mathematical morphology,and then calculated the crack length,and built a road detection application system to make the crack detection results more intuitive.In summary,the paper has completed the task of detecting road cracks,using deep learning algorithms to complete the paper’s detection work,providing new ideas for road maintenance,and providing road maintenance decision-making data support.