Research On Semantic Segmentation Of Road Cracks Integrating Dense Connection And Attention Mechanism

China is a strong country in infrastructure construction,and its highway mileage is constantly refreshed,ranking second in the world.However,the rapid development of infrastructure will inevitably bring heavy maintenance and repair work.Therefore,road maintenance has gradually become the focus of current traffic management.The research of road crack is an important task in road maintenance.Realize the statistics and calculation of road cracks,can find out the damage of the road surface in time,provide accurate data,and facilitate the relevant departments to carry out digital,scientific,and refined management.This paper uses image processing technology and semantic segmentation algorithm to achieve accurate segmentation of road cracks.This plays an important role in saving labor costs,improving the level of road maintenance,effectively implementing road maintenance,and ensuring driving safety.The study of road cracks in this paper is as follows:Firstly,the data set of road cracks is constructed.The data set contains three types of cracks,which are single cracks,cracked cracks,and repaired cracks,with a total of 8,400 pictures,laying the foundation for subsequent model building and training.After that,four semantic segmentation networks were trained and tested to evaluate the segmentation accuracy of each network on the data set.The DeepLabv3+ network has the highest segmentation accuracy on the test set,reaching 75.5% MIo U,and the average pixel accuracy is 92.1%.Therefore,this network is selected as the benchmark network in this paper.At the same time,the causes of the small cracks in the segmentation process are summarized and analyzed,such as the inaccuracy of the segmentation and the cavity of the segmentation.Then,in order to further improve the accuracy of road crack segmentation,the benchmark network is improved.In view of the complete segmentation of small cracks and holes in large-scale cracks,the dense connection mechanism of Dense Net is integrated into the ASPP module of the reference network.The spatial pyramid pooling module connects a group of convolutional layers with a dense connection to generate Multi-scale features can significantly increase the model’s receptive field and feature extraction capabilities.The experimental results show that the improved network can not only complete the segmentation of small cracks,but also effectively solve the problem of voids in the fracture segmentation.The MIo U of the improved network on the test data set reached 80.2%,an increase of 4.7% compared with the benchmark network.Finally,aiming at the pavement shadow in the process of crack segmentation,two types of attention modules are added on the basis of the improved network,which are spatial attention module and channel attention module respectively.The spatial attention module can calculate the correlation between each pixel and other pixels in the image,and the correlation between pixels does not change with the distance.This correlation can be used as weight to act on the feature map,strengthen the close connection between crack features and ensure the integrity of crack segmentation.The channel attention module can selectively express the feature,pool the feature map with advanced semantics,and then connect it with the full connection layer.Through supervised learning,the channel of the feature map is given different weights to enhance the expression of the crack feature.After integrating these two modules into the benchmark network selected in this article,reasonable experiments were carried out to improve the characteristic representation of the network,and finally the MIo U on the test set reached 83.9%.The algorithm proposed in this paper can be effectively applied to crack segmentation in road scenes.The algorithm has high accuracy and can be applied in practice to reduce the loss of manpower and material resources caused by manual statistics.