基于YOLO算法的探地雷达道路图像异常自动检测

探地雷达(ground-penetrating radar, GPR)是一种可用于道路内部异常目标识别的无损检测方法。GPR工作时往往产生海量的扫描数据，而数据解释是技术要求高、任务繁重的工作，通常需要人工完成。此外，道路内部的复杂性和异常目标的多样性增加了图像异常检测的难度。近年来，人工智能(artificial intelligence, AI)技术的快速发展为基于AI的探地雷达B-scan图像自动解释提供了可行的技术思路，常用的深度学习算法有RCNN(region-convolutional neural network)和YOLO(you only look once)。虽然YOLOv3在目标检测方面已经有了一定的成效，但YOLOv4的改进算法可以进一步提高检测能力。结合YOLOv3算法，对比研究分析YOLOv4目标检测算法的改进对于目标检测任务的影响，以及YOLOv4算法对探地雷达图像异常目标检测效率的提升能力。结果表明，YOLOv4的改进算法更适用于探地雷达异常目标的自动检测，经过训练后的YOLOv4网络模型满足探地雷达道路内部异常目标智能化检测需求，具有较强的实用价值。

Automatic Detection of Anomalies in GPR Images Based on YOLO Algorithm

Ground-penetrating radar (GPR) is a non-destructive testing method used for identifying internal anomalies in roads. GPR often generates a massive amount of scan data during operation, which requires high technical skills and a heavy workload for interpretation and is typically completed manually. Moreover, the complexity of road structures and the diversity of anomalies increase the difficulty of image anomaly detection. In recent years, the rapid development of artificial intelligence (AI) technology has provided feasible technical solutions for automatic interpretation of GPR B-scan images based on AI. Common deep learning algorithms for object detection include RCNN and YOLO. While YOLOv3 has achieved some success in object detection, the YOLOv4 algorithm based on its improvements can further enhance detection capabilities.This paper investigates the improve the ability of object detection. This paper investigates the impact of the improvement in YOLOv4 on the object detection and the ability of YOLOv4 improving the efficiency of detecting abnormal object in the GPR image. The results show that YOLOv4 with the network framework and algorithm improvements is more suitable for automatic detection of anomalies in GPR images. The trained YOLOv4 network meets the requirements of intelligent detection of abnormal object in GPR images and possesses the high practical value.