Multi-view Radar Map Road Disease Detection Algorithm Based On Deep Learning

With urbanization,transportation infrastructure is gradually being upgraded,and various roads are being developed based on geographical circumstances and usage.However,with extended use,natural environmental conditions and human causes both can create varied degrees of road diseases.These road flaws not only affect driving comfort,but they also pose a safety risk if they build over time and are not remedied promptly.As a result,rapid detection of road structural diseases is critical to extending road life and ensuring safe vehicle operation.By virtue of its ability to identify the internal states of pavement structures,threedimensional ground-penetrating radar(3D-GPR)is one of the common means of road inspection nowadays.The old manual approach of diagnosing road damage is inefficient and time-consuming due to the enormous number of roads and the complicated imaging of data collected by 3D-GPR.As a result,the current focus in the field of road radar disease detection is to use automatic detection as an alternative to manual detection.Although 3D-GPR captures information in different views,existing studies often only evaluate data in one direction,making full utilization of multi-view data difficult.Simultaneously,the training process of model parameters in deep learning algorithms frequently requires a considerable amount of labeled sample data,and the size of the sample data defines the performance of proposed model.The scarcity of radar maps with labels,on the other hand,substantially inhibits the progress of relevant research.As a result,the diverse input features of radar data and the shortage of annotation are significant obstacles encountered in the field of road detection study based on radar datasets.Therefore,after analyzing the four characteristics of multi-view input,unbalanced data distribution,abundant irrelevant interference information,and small sample size in radar mapping data,a two-stage model for multi-view road disease detection is proposed based on attention fusion and knowledge distillation in this paper.To make full use of multiple input features of multi-view radar data with a small number of samples,this paper proposes a dual-tower model based on attention fusion(Dual-AttFusioNet)with automatic feature extraction by deep learning to achieve automatic detection of road disease categories.First,the dataset is pre-processed and then balanced and expanded using up-sampling and data augmentation methods.Then,combining View-wised attention mechanism and Channel-wised attention mechanism to achieve efficient and accurate end-to-end detection and identification of road diseases.To further improve the practical application of the model in road inspection equipment,reduce the hardware cost and speed up the decision,the complex dual-tower model is distilled using knowledge distillation to train the lightweight densely connected network(DenseNet-63)and lightweight convolutional neural network(CNN-6)with a small number of parameters and shallow layers.Experimental results show that the model proposed in this paper can achieve high accuracy on both the original and augmented datasets using only fewer parameters and computations.