Research On Multi-scale Pavement Crack Detection Algorithm Based On Attention

The pavement to be used for a long time will be damaged.Pavement crack is one of the common distress of pavement directly threatens the safety of vehicles and service life of the highway.With the rapid development of China’s transportation system,road maintenance plays an increasingly important role in the intelligent transportation system.Traditional manual detection of pavement crack consume a lot of manpower and material resources.At the same time,these approaches are very time-consuming,dangerous,labor-intensive and subjective,which can not the development requirements of China’s transportation system.So it is necessary to design an automatic,accurate and efficient pavement crack detection method in the field of intelligent transportation.Recently,with the rapid development of information technology,especially the rapid development of image processing technology and deep learning technology,the application of computer vision technology instead of the traditional artificial method for pavement crack detection has attracted more and more attention.Pavement crack detection is challenging due to the low contrast between cracks and the surrounding pavement,the intense inhomogeneity along the cracks and topology complexity of cracks.In this paper,the deep learning method of convolutional neural network is introduced into the detection of road crack.This paper proposes a novel pavement crack detection method based on an end-to-end trainable deep convolution neural network which can detect the pavement crack at the pixel level.We build the network using the encoder-decoder architecture and adopt the improved pyramid module to complete the multi-scale detection,which can extract crack information from a global view and enhance the continuity of pavement crack detection.Moreover,we introduce a spatial-channel combinational attention module into the encoder-decoder network for refining crack features and improve the representation of the network.In addition,the existing method are probably loosing crack details due to the reduction of feature map resolution,which is fatal for the prediction of small objects.Therefore,the dilated convolution is used to reduce the loss of crack details.Moreover,after detecting the pavement crack at the pixel level,by referring to the existing classification methods of road cracks,we designed the classification algorithm of pavement cracks which can divide pavement crack into transverse cracks,longitudinal cracks,and crevasse cracks.In addition,the characteristic information of crack width and length are measured.In order to prove the superiority of this method,comparative experiments have been completed on three public datasets.From the experimental results,it can be seen that this method can detect crack more accurately,effectively and quickly than existing pavement crack detection methods,and the detection results are consistent with human visual characteristics.