Adaptive Curriculum Learning for Enhanced Detection of Subsurface Distress Using Ground Penetrating Radar
Deep learning combined with ground penetrating radar(GPR)imaging has become a popular research area for detecting subsurface distress in roads.However,due to the difficulty in obtaining ground truth samples,improving the accuracy of subsurface distress detection under small-sample conditions remains challenging.To address this issue,this paper proposes a self-adaptive curriculum learning model framework to enhance the detection accuracy of subsurface distress.The framework follows a progressive training strategy from easy to difficult,which can dynamically evaluate the difficulty of sample data and adaptively adjust the training order of data.Firstly,a teacher-student adaptive curriculum learning framework is constructed,which alternately optimizes the sample difficulty evaluation mechanism of the teacher network and subsurface distress detection performance of the student network.A soft-boundary constrained loss function is introduced to dynamically evaluate the difficulty of samples based on the prediction error of the detection model,and the adaptive scheduling of sample training order is realized by the backpropagation of weighted loss.To adapt to different scales and types of detection models,the EfficientNet series network is used to construct the feature extraction module.Based on actual engineering data,a dataset containing three types of road subsurface distress-reflective cracks,loose,and voids-is constructed with a total of 1 857 samples.Four strategies are designed in the experiment:adaptive curriculum learning,random training,specified order,and specified reverse order,which are verified by two subsurface distress detection algorithms,YOLOv9 and DETR.The results show that starting training from simple data samples and gradually increasing the difficulty of samples can effectively accelerate model convergence and improve detection accuracy.Compared with the other three methods,the proposed adaptive curriculum learning framework improved the detection accuracy of subsurface distress by 6.21％,4.01％,and 10.64％,respectively.By comparing the difficulty evaluation of each sample,it is found that the detection difficulty of subsurface distress in urban road scenes is greater,but there is no significant difference in the difficulty level between different subsurface distress types.This study provides a new technical approach for the detection of road subsurface distress,which will help to improve the accuracy and efficiency of detection.
NETL
NSTL
万方数据
