Research On Concrete Pavement Crack Detection Algorithm With Optimized Side-Output Fusion And Attention Mechanism

The rapid development of the economy promotes the improvement of people’s quality of life,which indirectly promotes the high-frequency use of roads and the increase of road mileage year by year.The country has introduced various policies to support the completion of intelligent transportation in order to build a strong transportation country,in which high-quality roads play an important role.The huge road mileage leads to a large amount of work in highway maintenance,and the focus of maintenance is the damage to pavement cracks.Cracks should be detected early,the severity assessed,and corresponding maintenance decisions made.The earlier crack detection work mainly relied on manual detection with low efficiency and low return rate.Now,the need for automated crack detection is imminent in the face of such a huge workload.Deep learning has made great achievements in all walks of life,and it is also widely used in the transportation industry.In the field of crack detection,various network models have been proposed,and the development is accompanied by difficult problems that need to be broken.To address the shortcomings of crack detection algorithms,the main research work of this thesis includes:1.Aiming at the common problems of low accuracy of crack detection results and sensitivity to noise,inspired by the latest development of crack detection at home and abroad,a concrete pavement crack detection algorithm of hierarchical feature fusion and connected attention architecture is proposed in this thesis.The deeper network structure,Res Net50,can better remove the influence of noise,and can detect crack pixels more accurately with the enhanced attention module.In the side-output network,the features of the low and middle layers are fused by different modules to maximize the advantages of each side feature.2.Aiming at the problem of missing details and low resolution due to the loss of crack details when the image size is first pooled and reduced in the convolution process,and then expanded by deconvolution,a fusion module of parallel lightweight dilated convolution is used in the side-output network to combine the middle-level and high-level features.The parallel dilated convolution not only changes the limitation of the receptive field,but also aggregates more effective feature information.The lightweight dilated convolution effectively controls the parameters of the network,so that the addition of the dilated module can reduce the impact of the detection speed of the model as much as possible.3.Aiming at the problem that the resolution is gradually reduced during the convolution process and the positive and negative samples of the crack image are unbalanced,the idea of progressively and repeatedly polishing the features is proposed,and the weighted loss function combined with the deep supervision training strategy is used to help the model converge better.The features of the low-level and middle-level are fused into the backbone network and polished again.On the one hand,the resolution of the high-level input is improved,and on the other hand,the feature information is increased and the noise is reduced.In the side-output network,the lightweight dilated convolution module is used to polish the middle-level and high-level feature information again to provide strong support for the final prediction map.Finally,a systematic summary of the research work in this thesis is described and an outlook on the future challenges that need to be continued to be tackled is given.