探地雷达图像异常检测方法研究及应用

Research and Application of Anomaly Detection Method for GPR Image

探地雷达技术广泛应用于城市道路病害探测领域。道路异常检测通常需要人工目视解译来实现,存在较强的主观性和不确定性。本文基于YOLOv5目标检测算法进行设计改进,提出了一种探地雷达图像道路空洞异常检测的方法。该方法添加了一个微小目标特征检测的预测头,并引入卷积块注意力模块和高效的交并比损失函数。模型不仅迭代收敛快、回归精度高,还优化了边界框回归任务的样本不均衡问题。通过消融实验表明,改进后的方法在道路空洞检测应用中检测精度明显得到提升,平均精度均值由81.62%提升到83.90%,对道路空洞病害异常的检测效果有很大提升。

Ground penetrating radar(GPR) technology is widely used in the field of urban road disease detection.Road anomaly detection usually requires manual visual interpretation,which has strong subjectivity and uncertainty.In this paper,based on YOLOv5 target detection algorithm design improvement,proposed a GPR image road hole anomaly detection method.This method adds a prediction head for tiny target feature detection,and introduces CBAM and EIoU loss function.The model not only has fast iterative convergence and high regression accuracy,but also optimizes the sample imbalance problem of the bounding box regression task.The ablation experiment shows that the detection accuracy of the improved method is significantly improved in the application of road cavity detection,and the average accuracy is increased from 81.62% to 83.90%,which greatly improves the detection effect of road void disease abnormality.