Automatic End-to-end Road Crack Detection Using Deep Learning

With the advancement of socialist modernization,China’s infrastructure construction is booming.The mileage of expressway,as a symbol of national modernization,has leaped to the first place in the world in recent years.However,such a huge and complex road traffic network also brings problems like road maintenance.Because of construction mode,climate conditions,traffic load,road crack often occurs,which brings significant challenges to traffic safety.Therefore,how to detect crack area accurately from the road image has become an urgent problem to be solved.Traditional road crack detection methods are easily disturbed by noise and other factors,even with the road material changes,the result of detection will be greatly reduced.In recent years,convolution neural network has made exciting achievements in image semantic segmentation.Therefore,it is become a worthy research direction that applies convolution neural network to road crack detection.Convolutional neural networks usually need a large number of labeled data to train,but current data volume cannot meet the demand.Therefore,in this paper,we first apply a variety of image transformation methods to expand datasets.Then,in order to solve the problem of large similarity of the amplified data and further expand the datasets,a data synthesis method is proposed.Firstly,the crack areas are extracted from the image as foreground.Then after transformation,we combine the foreground with the background image that without crack areas using poisson fusion.We can synthesize enough labeled data,which meets our demand for large-scale data.In this paper,a deep convolution neural network with decoder-encoder structure is designed for automatic end-to-end road crack detection.The network consists of 108 convolution layers,which are composed of down sampling path,bottleneck and up sampling path.The down sampling path gradually reduces the size of feature maps to extract image features;the bottleneck is used to reduce the number of feature maps and eliminate redundant information;the up sampling path gradually expands the size of feature maps to recover image details and predict target labels.Through a large number of experiments and comparisons with other related methods,our method has achieved the state-of-the-art results in several evaluation metrics,and significantly surpass other methods,which proves effectiveness of the method in this paper.