Research On Semantic Segmentation Of Pavement Crack Based On Deep Learning

By 2021,the total length of roads in China has exceeded 5 million kilometers,ranking first in the world.Maintaining good road conditions is an important responsibility of national and local road maintenance departments.Due to long-term weathering,precipitation,stress,and other factors,different types of cracks appear on the surface and inside structure of road pavement,which can cause traffic accidents.Therefore,timely and accurate crack detection is crucial for highway maintenance and accident prevention.Currently,road surface inspections mainly rely on manual inspection or a few intelligent algorithms.Manual inspection methods completely dependon the knowledge and experienceof engineering inspectors,making ithighly subjective,while existing intelligent algorithms have low accuracy.Consequently,how to effectively detect and accurately obtain pavement crack information has become a focal point of currentresearch.In recent years,as a research topic in the field of computer vision,semantic segmentation technology has played an important role in real-world scenarios such as intelligent security,agricultural production,medical diagnosis,and urban planning and management.With the continuous development of deep learning technology,the performance of image semantic segmentation based on deep learning has greatly improved,with the advantages of automatic feature extraction,high recognition accuracy,and strong noise resistance.Therefore,applying depth learning based image semantic segmentation technology to the field of pavement cracks can provide assistance in accurately obtaining pavement crack information.However,the existing semantic segmentation methods for pavement cracks based on deep learning still need toimprovetheirsegmentation accuracyinterms ofcomplexpavement crack backgrounds,fuzzy boundaries,and small crack segmentation.In response to the above issues,the research content ofthis articleisas follows:(1)Aiming at the complex background of pavement crack,an improved semantic segmentation method for pavement cracks based on U-Net network is proposed.This method can enhance the identification of crack regions by integrating attention mechanisms into the hop connections of U-Net networks;At the same time,introducing extended convolution into the encoder decoder architecture can reduce the loss of crack details caused by pooling operations in theencoder network.Inaddition,due to theloss ofcrack information in thedecoder network due to upsampling,this thesis proposes a depth separable residual module to capture road crack information.The experimental results show that the model evaluation index of this method on two complex background pavement crack datasets is superior to current mainstream methods,whichprovestheeffectivenessofthismethod.(2)Aiming at the problem of fuzzy boundary of pavement cracks,an improved model based on Seg Net network is proposed to segment pavement cracks.This method designs a new up sampling and hop connection process between the encoder and decoder networks,which can enhance feature fusion and information exchange at adjacent levels through the reuse of feature maps,and preserve pavement crack boundary information.At the same time,combining attention mechanism and series parallel expansion convolution in an encoder decoder network can improve the accuracy of crack boundary location.This thesis proves the effectiveness of this methodontwo fuzzydatasetsoffractureboundaries.(3)To solve the problem of discontinuous segmentation of small cracks in the pavement,this thesis uses a GAN-MUNet based method for semantic segmentation of pavement cracks.Due to the characteristics of the adversarial generation network that requires less training data,produces good results,and makes the segmentation results more continuous,this thesis combines the adversarial neural network and the semantic segmentation network.Experimental results show that this method effectively improves the problem ofdiscontinuous segmentation of smallpavementcracks.Figure[45]Table[11] Reference[61]...