Research On Pavement Damage Detection Method Based On Deep Learning

Pavement crack detection is of great importance for safe driving.There are two main problems with existing pavement crack detection networks.The first problem is the poor performance of existing pavement crack detection methods in the presence of large variations in crack scales.In addition,these networks lack the ability to strengthen the crack features and weaken the background features,resulting in the networks’ inability to exclude the interference of background information on crack feature extraction.The second problem is that the existing pavement damage detection method ignores the problem that there is a large difference in the target width to height ratio for this detection task,resulting in a less than reasonable distribution of positive and negative samples.At the same time,the suppression of background information is still insufficient,and the anti-interference ability of model background information needs to be further improved.Aiming at the first problem,a novel pavement crack detection network SDFNet(Scale adaptation,Deep layer and shallow layer attentional feature fusion,Feature refinement Network)is proposed in this thesis.In order to fully extract the multi-scale feature information of cracks,a novel scale adaptive module is proposed.The module obtains deformable damage features with adaptive scaling by using deformable convolutional tandem operations with a suitable combination strategy of expansion coefficients and suitable jump connections.In order to strengthen the discriminative power of the model for the damaged region and alleviate the interference of the background region,a novel deep layer and shallow layer attentional feature fusion module is proposed.This module captures the channel information in the deeper layers of the network by channel attention,and uses the proposed shallow feature extractor to obtain shallow features,which are then fed into the proposed chunked decoupled attention mechanism to fuse shallow spatial information with deep spatial information to obtain spatial information.After that,the channel information is fused with spatial information,and finally the proposed feature refinement module strengthens the fracture features and weakens the background information.Aiming at the second problem,this thesis proposes a pavement damage detection method based on an improved dynamic label assignment scheme and a traditional feature enhancement network.The improved dynamic label assignment scheme takes into account the aspect ratio of the anchor frame and the real frame by refining the regression cost function of Sim OTA(Similarity Overlap Threshold Assigner)to calculate the cost matrix,thus making the network more reasonable for positive and negative sample assignment in the pavement damage detection task.In addition,a traditional feature enhancer is proposed for the background interference problem.This module extracts the background information based on a simple adaptive thresholding method,and obtains the suppression information on the background region through processing,and fuses it with the features extracted by the backbone network to achieve the suppression of the background information.In this thesis,the above method is fully experimented.The experimental results show that the multiscale pavement crack detection method based on the deep layer and shallow layer feature attention mechanism proposed in this thesis can effectively obtain the multiscale information of the image and can effectively enhance the ability of the model to extract the features of the damage area.In addition,the pavement damage detection method based on the improved dynamic label assignment scheme and the traditional feature enhancer proposed in this thesis can not only alleviate the influence of background information on the model feature extraction,but also make the positive and negative sample assignment of the model more reasonable in the training process.The proposed method also outperforms the existing mainstream methods on the publicly available datasets RDD2020 dataset and CNRDD dataset.