测绘通报 ›› 2026, Vol. 0 ›› Issue (8): 154-160,173.doi: 10.13474/j.cnki.11-2246.2026.0822

• 技术交流 • 上一篇    

基于YOLO11改进网络的探地雷达地下病害智能识别

黄彬1, 石国杰1, 杜力1, 王林柯1, 韩沐阳1, 彭鸿伟1, 李克凡2, 李佳玺2, 王晋国2, 侯兆阳2   

  1. 1. 联检(江苏)科技股份有限公司, 江苏 常州 213015;
    2. 长安大学理学院, 陕西 西安 710000
  • 收稿日期:2025-11-17 发布日期:2026-09-12
  • 通讯作者: 侯兆阳。E-mail: zhaoyanghou@163.com
  • 作者简介:黄彬(1979—),男,硕士,工程师,主要从事道路桥梁和建筑房屋的检测和研究。E-mail: huangbin@czjky.com
  • 基金资助:
    国家自然科学基金(52401168)

Intelligent detection of subsurface defects in ground penetrating radar images based on an improved YOLO11 network

Huang Bin1, Shi Guojie1, Du Li1, Wang Linke1, Han Muyang1, Peng Hongwei1, Li Kefan2, Li Jiaxi2, Wang Jinguo2, Hou Zhaoyang2   

  1. 1. Lianjian (Jiangsu)Technology Co., Ltd., Changzhou 213015, China;
    2. School of Science, Chang'an University, Xi'an 710000, China
  • Received:2025-11-17 Published:2026-09-12

摘要: [目的] 本文旨在解决探地雷达(GPR)B-scan图像在人工智能识别中存在的噪声干扰强、目标边缘模糊及模型复杂度高等关键问题。[方法] 本文提出了一种基于改进YOLO11的轻量化地下病害检测方法——YOLO11_DA。该方法在网络结构中融合动态卷积(C3k2_DynamicConv)与自适应下采样(ADown)模块,以增强特征提取能力并保留关键低频信息。[结果] 消融试验结果表明,改进后模型在准确率、召回率、mAP0.5与mAP0.5:0.954项指标上分别达到91.3%、86.1%、87.7%和58.8%,参数量降至2.61×106,推理速度达 429.56 帧/s。[结论] 相较于YOLOv5、YOLOv8、YOLOv10及基准YOLO11模型,本文方法在保持轻量化的同时,在复杂噪声、弱回波与多尺度场景下表现出更优的稳健性与泛化能力,为道路地下病害的智能检测提供了有效的技术解决方案。

关键词: 道路检测, 探地雷达, 深度学习, YOLO11, 3Ck2_DynamicConv, ADown

Abstract: [Purposes] This paper aims to address the strong noise interference,blurred target edges,and high model complexity in the intelligent detection of ground penetrating radar (GPR)B-scan images. [Methods] This work proposed a lightweight underground defect detection method based on an improved YOLO11 framework.This approach integrates a dynamic convolution and adaptive down-sampling modules into the network architecture of YOLO11 to enhance feature extraction capability while preserving critical low-frequency information. [Findings] The experimental results demonstrate that the improved model of YOLO11_DA achieves a precision of 91.3%,recall of 86.1%,mAP 0.5 of 87.7%,and mAP 0.5:0.95 of 58.8%,while reducing parameters to 2.61×106 and achieving an inference speed of 429.56 frames per second. [Conclusions] Compared to benchmark models including YOLOv5,YOLOv8,YOLOv10,and the original YOLO11,the proposed method maintains lightweight characteristics while exhibiting superior robustness and generalization capability in complex noise environments,weak echo conditions,and multi-scale target scenarios,which has provided an effective technical solution for intelligent detection of underground road defects.

Key words: underground road detection, ground penetrating radar, deep learning, YOLO11, C3k2_DynamicConv, ADown

中图分类号: