With the development of global highway construction,the total mileage of the world’s roads increases rapidly,a large number of old roadbed began to appear a variety of hidden defects,to the road traffic safety brought huge hidden dangers.Therefore,the worldwide demand for road condition detection and maintenance has increased dramatically.Ground penetrating radar(GPR)is gradually replacing the impact echo method,ultrasonic method,drill core sampling method and rebound method as the main means of testing civil infrastructure such as roads,bridges and tunnels due to its fast detection speed,large detectable depth and no contact.However,the texture information in the image data generated by GPR mainly depends on interpretation by professionals,which not only has low interpretation efficiency,but also the accuracy of interpretation varies from person to person,which restricts the large-scale promotion of GPR detection methods.In order to solve this problem,this paper proposes to use deep neural network technology to build a highway subgrade internal defect detection system with artificial intelligence,which can realize automatic identification an d location of roadbed defects.The framework of highway subgrade internal defect detection system developed in this paper includes three parts.The first is the formation of database,which is composed of simulation,road acquisition,test block test and data enhancemen t.Secondly,the framework has the function of preprocessing data and generating simulation files in batches.Finally,the framework platform can provide the choice of classical detection model or self-training model to diagnose the input data and distinguish the location and classification of the disease.In this paper,we propose an aggregated decentralized downsampling algorithm for the imaging characteristics of radar echograms,which re-extracts features from multiple feature maps that would otherwise be lost and fuses them together to form richer feature information,making the feature extraction capability of the whole network nearly enhanced.The network with enhanced feature extraction capability achieves up to 2.65% improvement in accuracy on t he homemade dataset compared to the unimproved network.To address the potential pitfalls of aggregated decentralized downsampling algorithm in the fusion process that may cause redundant information as well as generate misleading information,a neural net work algorithm based on multi-region convolutional stacking is further proposed to further improve the recognition rate of grayscale maps transformed from multiple invisible disease radar data.The accuracy of the 34-layer network using multi-region convolutional stack reaches 96.97%,which is 3.95% higher than the original 34-layer ResNet network.Finally,this paper analyzes the proposed algorithm and summarizes the functions of the detection system.In addition,it also discusses the space for further improvement and research of the proposed algorithm. |