Research On Algorithm Of Crack Detection Based On Multi-scale Dilated Convolution And Dual Attention Mechanism

As the foundation of transportation,the highway plays an important role in the development of the national economy.Due to overload,improper use,long-term rain corrosion,and other problems,some road sections are damaged to different degrees,such as cracking and peeling.Pavement cracks are usually used to measure the quality of roads.If early cracks can be found in time and repaired and tracked,the consumption of manpower,material resources,and financial resources will be greatly reduced.Therefore,pavement crack detection has become an important research topic at present,and timely detection and repair of pavement cracks is also an important part of road maintenance.The traditional crack detection method based on digital image processing has a weak denoising ability and poor generalization effect for crack images with loud noise and complex topological structure.To solve this problem,this thesis uses deep learning method to identify pavement cracks and proposes a pavement crack detection method based on multi-scale dilated convolution and dual attention mechanism by using the rich hierarchical features of Convolution Neural Network(CNN).The main contents and innovations of this thesis are as follows:1.Aiming at the problem that the pavement crack detection method based on digital image processing has low accuracy and poor generalization performance,in the detection process for crack images with complex topological structure and great influence by noise,this thesis studies the crack detection method based on deep learning,and uses the depth residual network Res Net-101 to extract features.At the same time,in order to extract more meaningful features,this thesis introduces a dual attention model into the backbone network and embeds the attention mechanism into Res Net-101 to enhance the network model’s focus on crack pixels.2.Aiming at the problem that a large amount of detailed information is lost when the network deepens the resolution of the feature map,this thesis studies the method of multi-scale dilated convolution to increase the receptive field of convolution kernel without reducing the resolution of the feature map,so as to improve the positioning ability of the network model to crack pixels.3.Aiming at the problems of large receptive field of deep network,strong semantic representation ability,high resolution of shallow network feature map and abundant detail information,the feature pyramid fusion method is adopted to fully integrate deep semantic information layer by layer with shallow detail information to extract rich crack features.Finally,the edge fusion network is used to integrate the extracted features at all levels to obtain the final crack prediction image.This thesis is trained and tested on four datasets.The experimental results show that the improved convolutional neural network model has a good ability to identify fine cracks,wide cracks,and cracks with complex topological structures on concrete pavement and asphalt pavement,which is 9.9% higher than the original Res Net-101 model in F-score value,proving that the network model proposed in this paper has a good generalization ability.