Computer Vision Based Research On Crack Detection Of Road Pavement

With the development of national infrastructure,the roads built by concrete and asphalt are all over the country.With the daily traffic of vehicles,there are inevitable cracks on the road surface,not only endangering the safety of driving,but also facing the risk of collapse.Therefore,the detection and monitoring of road safety is particularly important.Traditional manual detection methods have problems such as low efficiency and low accuracy.Therefore,the goal of this thesis is to study the algorithms for automatic crack detection,marking and area size calculation based on the road photos collected by the machine,and finally package it into application software.In this thesis,set of automatic detection system of road cracks is proposed.In the algorithm part,two ideas are tried.The first idea is to combine convolution classification network and digital image processing related algorithm,cut the original image into small pieces,then use Inception V3 network to complete the classification of small pieces,and use digital image processing related algorithm to complete the crack extraction after the classification.In the convolution classification network,this thesis makes a comparative experiment to optimize the batch size,learning rate,network structure,loss function,etc.According to the uneven distribution of the data set,a weighted cross-entropy loss function is used.The network is reduced in speed due to the deep and excessive redundancy of the network,and the network is reduced under the condition of reducing the accuracy loss.In the aspect of digital image processing,some algorithms are used,such as mask difference smoothing,gray-scale stretching,median filter drying,OTSU threshold segmentation,corrosion expansion connecting connected domain,based on the crack shape drying,crack skeleton refinement,crack skeleton burr removal,etc.,and a variety of algorithms are combined to complete the segmentation and extraction of cracks.The second idea is to directly extract the crack region based on the full convolutional network based on semantic segmentation.In this thesis,VGG16 was selected as the coding structure of the full convolutional network through control experiments,and UNet as the overall architecture of the full convolutional network.According to the characteristics of the small area of the crackarea,a weighted summation of the cross-entropy loss function and the IOU loss function of the crack area is proposed to improve the loss function,and a convolution layer is added at the skip connection to optimize the network structure.In addition,this thesis does cross-validation of the proposed algorithm on two different data sets,and compares it with other crack algorithms in other papers.At the end of this paper,the two algorithms are accurately compared,and the results show that in the field of road crack recognition,the full convolutional network has a better recognition effect on crack features than traditional algorithms,and has certain application value in the field of road crack automatic recognition.Finally,a full convolutional network is selected here to complete the software for automatically identifying cracks,which can be provided to engineering personnel for automated processing of crack data in the professional field.