Research On Pavement Crack Segmentation Method Based On U-Net

The highway is an important infrastructure in our country and plays a vital role in the national economy.The rolling of various heavy vehicles,the change of ambient temperature and the infiltration of water and other reasons lead to different degrees of pavement damage,among which crack is the initial state of all damage.Early pavement crack detection is mainly done by manual work,but manual detection shows low efficiency,low precision,strong subjectivity and other limitations.So it is necessary to realize automatic pavement crack detection.In the past period of time,people first used handmade features such as pixel intensity,edge and texture to detect pavement cracks.The commonly used methods include threshold method,edge detection method and region growth method.These handmade features are low-level representations and less robust to stains,lighting conditions and other disturbances.Later,influenced by the powerful feature acquisition and expression ability of deep convolutional neural networks,researchers developed a large number of crack detection methods based on deep convolutional neural networks,but some existing methods still have the problem of inaccurate segmentation of some small and slender cracks.Therefore,based on U-Net network,the paper studies and improves the existing problems in the field of pavement crack detection.The main research contents of the paper are as follows:(1)In view of the phenomenon of resolution loss caused by pooling operation in feature extraction part of U-Net network,a Hybrid Dilated Convolution(Hybrid Dilated Convolution,HDC)was introduced into the skip connection position of U-Net network to alleviate the problem of detail feature loss caused by resolution loss during sampling on U-Net network.HDC is a hybrid dilated convolution block that can expand the receptive field of the network without loss of spatial resolution,thus enhancing the ability of the network to acquire global features and long-distance information.The network is named the HDCU-Net network.Experiments are designed to compare HDCU-Net algorithm with other existing methods on a public dataset.The experimental result shows that the proposed HDCU-Net network can extract more detailed and complete cracks.(2)In order to solve the problem that HDCU-Net cannot give more attention to the crack region,a module called Hybrid Attention Mechanism(Hybrid Attention Mechanism,HAM)is proposed to be added to the skip connection location of the HDCU-Net network,so that the network can focus more attention on the crack location of the input image.HAM is a kind of module inspired by CBAM,which combines the channel attention mechanism and the spatial attention mechanism.This module can improve the network’s attention to the cracked part,and thus improve the accuracy and generalization ability of this network model.The network is called HADCU-Net.The HADCU-Net network and other existing algorithms are tested on multiple datasets,and the experimental results show that HADCUNet has further improved the evaluation criteria in the pavement crack segmentation task.(3)In order to solve the problem that most encoder-decoder structure networks do not pay much attention to the feature fusion of decoder,the Multiscale Decoder(Multiscale Decoder,MD)and Residual Connection(Residual Connection,RC)mechanism are introduced based on HADCU-Net.MD is to add deconvolution operation to the original decoder part in parallel,while RC is to improve the connection mechanism in the decoder.The above two operations are to optimize the multiscale feature fusion.The network is named HMADCUNet.The proposed algorithm is compared with many other existing algorithms on open datasets,and the experimental results show that this proposed network has better performance than other algorithms.