Research On Image Crack Recognition Of Concrete Pavement Based On U-net Neural Network

With the continuous development of highway transportation,our country has gradually become a major transportation country.The mileage of roads has been increasing,and road maintenance has gradually become a research hotspot.Cracks are an early manifestation of concrete pavement diseases.Early detection and early treatment of cracks play an important role in the maintenance of the pavement.Compared with the traditional manual detection method,the automated crack detection technology has the advantages of low cost and high efficiency.With the continuous breakthrough of computer hardware technology,deep learning algorithms have been continuously optimized.Compared with algorithms based on traditional digital image processing,they have the advantages of being more accurate and robust.Therefore,the research on crack image of concrete pavement has become a trend.The crack detection algorithm of concrete pavement is studied in this thesis.Aiming at the problem that the current crack detection algorithm has low recall of cracks that caused by noise,an improved U-net crack detection algorithm is proposed.In order to further improve the accuracy of the crack detection algorithm,a full U-shaped network structure is proposed to extract cracks in the concrete pavement based on the U-net.The main research work of this thesis as follows:1.Aiming at the problem of the low recall rate that caused by the surface noise of the crack image of the concrete pavement based on the digital image processing algorithm,the U-net model was studied in depth,and its improved model was used for crack detection.First,in order to obtain more detailed information about the crack image,a down-sampling and an up-sampling process are added.Secondly,the original U-net model has reduced image specifications during convolution and lost edge information at the same time.In this thesis,the zero-fill method is used to ensure that the image size does not change during each convolution and to keep the edge information is complete.Finally,in order to increase the generalization of the network structure,a Dropout layer is added.Experimental results show that the improved U-net network structure can completely extract crack pixels in various environments.2.In order to further improve the crack detection accuracy of concrete pavement,a full U-shaped network structure is proposed based on the improved U-net structure in this thesis.First,the network is constructed based on the improved U-net model.Then,after each pooling layer,an upsampling is performed to restore its feature map size before the pooling layer and fuse with the convolution layer before pooling.The feature map after fusion is used as a new fusion layer to fuse with the network layer after network upsampling.After multiple feature fusions,global information is added while improving the network structure's learning of crack detail information.3.Investigate the maximum inter-class variance method to distinguish crack pixels from background pixels as completely as possible.At the same time,in order to eliminate as much as possible,the noise that erroneously recognizes the background pixels as crack pixels,and improve the precision of crack detection,a noise cancellation algorithm based on the characteristics of the connected area is studied.