Research On Automatic Interpretation Technology Of Ground Penetrating Radar Road Abnormal Targets Based On Deep Learning

Ground Penetrating Radar(GPR)is a non-destructive testing method used for identifying abnormal targets within the road.The interpretation of GPR data requires high technical expertise,real-time analysis,and involves a heavy workload.Typically,it is manually performed,placing demanding requirements on personnel in terms of technical skills,physical endurance,and mental focus.The complexity of the road’s interior results in complex backgrounds in GPR images,and the diverse forms of abnormal targets such as pipes,voids,and fractures further increase the difficulty of interpretation.In recent years,with the rapid development of Artificial Intelligence(AI)technology,it has provided a viable approach for efficient and automated interpretation of GPR road data based on AI algorithms.This study provides a comparative analysis of road anomaly target identification methods,including manual interpretation,traditional algorithm techniques,and deep learning methods.Drawing upon the superior achievements of deep learning algorithms in image recognition,the research focuses on exploring methods suitable for automatic identification of internal anomalies in road using GPR.The main content is outlined as follows:(1)Research on Road Anomaly Dataset for GPR.Firstly,different types of road models with anomalies are constructed,and gpr Max simulation software is used to perform forward simulations on these road models.The imaging characteristics of simulated images of anomalies at different locations and scales are analyzed and summarized to provide a theoretical basis for creating a simulated image dataset.Secondly,the types of internal anomalies in road structures detected by ground-penetrating radar and the characteristics of the corresponding measured data are summarized and analyzed.Finally,a comparative analysis is conducted to examine the differences between the simulated and measured data from ground-penetrating radar.(2)Automatic Classification of Road Anomalies in GPR using Deep Convolutional Neural Networks.Given that deep convolutional neural networks are more capable of extracting high-level abstract features,a deep convolutional neural network classification model was constructed for ground-penetrating radar data to achieve automatic classification of anomalies in a simulated dataset of road anomalies detected by ground-penetrating radar.(3)Automatic Recognition of Road Anomalies using GPR.In-depth research is conducted on the YOLO series deep learning algorithms for anomaly recognition,with a particular focus on comparative analysis and in-depth study of the YOLOv4 algorithm and YOLOv3 algorithm.This is done to investigate the impact of improving the YOLOv4 algorithm on the automatic interpretation of road anomalies detected by ground-penetrating radar and ultimately achieve automatic recognition of anomalies in real-world road data obtained from ground-penetrating radar.The results show that the automatic classification model based on convolutional neural networks achieved extremely high accuracy in the self-built simulated dataset for ground-penetrating radar.The automatic recognition model based on deep learning algorithms obtained an average accuracy of 93.52% and an average recall rate of 90.29%in the self-built ground-penetrating radar measured dataset.