面向探地雷达常见地下目标的GDS-YOLOv8n检测方法

王惠琴; 杨发东; 何永强; 刘宾灿; 刘鑫

doi:10.12000/JR24160

面向探地雷达常见地下目标的GDS-YOLOv8n检测方法

DOI: 10.12000/JR24160

王惠琴^1, ,,
杨发东^1, ,,
何永强²,
刘宾灿³,
刘鑫¹

1.
兰州理工大学计算机与通信学院兰州 730050
2.
西北民族大学土木工程学院兰州 730030
3.
陕西建工安装集团有限公司西安 710068

基金项目: 甘肃省重点研发计划 (23YFFA0060)，甘肃省优秀研究生“创新之星”项目(23YFFA0060)

详细信息

作者简介:
王惠琴，博士，教授，博士生导师，主要研究方向为雷达信号检测与处理，无线光通信理论与技术研究等

杨发东，硕士生，主要研究方向为雷达信号处理和深度学习

何永强，博士，教授，硕士生导师，主要研究方向为探地雷达信号智能处理，地质灾害分析与防治技术等

刘宾灿，高级工程师，主要研究方向隔震建筑的监测与性能评估、建筑节能及碳排放核算等

刘　鑫，硕士生，主要研究方向为探地雷达信号的去噪和反演

通讯作者:
王惠琴 whq1222@lut.edu.cn

杨发东 1948374001@qq.com

责任主编：雷文太 Corresponding Editor: LEI Wentai
中图分类号: TN957; P631
计量
- 文章访问数: 15
- HTML全文浏览量: 10
- PDF下载量: 1
- 被引次数: 0
出版历程
- 收稿日期: 2024-08-09
- 修回日期: 2024-10-14

Detection of Common Underground Targets in Ground Penetrating Radar Images using the GDS-YOLOv8n model

WANG Huiqin^{1
, ,},
YANG Fadong^{1
, ,},
HE Yongqiang²,
LIU Bincan³,
LIU Xin¹

1.
School of Computer and Communication, Lanzhou University of Technology, Lanzhou 730050, China
2.
School of Civil Engineering, Northwest Minzu University, Lanzhou 730030, China
3.
SCEGC Installation Group Company LTD, Xi’an 710068, China

Funds: Key Research and Development Program of Gansu Province (23YFFA0060), “Innovation Star” Project of Outstanding Graduate Students in Gansu Province

More Information

Corresponding author: WANG Huiqin, whq1222@lut.edu.cn; YANG Fadong, 1948374001@qq.com

摘要

摘要: 针对当前探地雷达(GPR)图像检测中存在准确率低、误检和漏检等问题，该文提出了一种GPR常见地下目标检测模型GDS-YOLOv8n。该模型首先使用DRRB特征提取模块替换YOLOv8n模型中的部分C2f模块，旨在增强模型对多尺度特征的提取能力。其次使用SPD-Conv下采样模块替换像素为320×320及以下特征图所对应的Conv模块，有效克服分辨率受限以及存在小目标的GPR图像在下采样过程中的信息损失问题；同时利用辅助训练模块，在不增加检测阶段模型复杂度的前提下提升GPR图像的检测性能。最后，引入Inner-SIoU损失函数，在添加新约束条件的基础上，通过比例因子生成适合于当前GPR图像的辅助边界框，以提高预测框的准确性。实验结果表明，GDS-YOLOv8n模型对金属管、PVC管和电缆线等6类常见地下目标在实测GPR图像数据集上的P, R和mAP50分别为97.1%, 96.2%和96.9%，较YOLOv8n模型分别提高了4.0%, 6.1%和4.1%，尤其对PVC管和电缆线目标的检测效果提升更明显。与YOLOv5n, YOLOv7-tiny和SSD等模型相比，其mAP50分别提高了7.20%, 5.70%和14.48%。此外，将GDS-YOLOv8n模型部署到NVIDIA Jetson Orin NX嵌入式设备上，并使用TensorRT进行优化。经FP16量化后，模型的检测速度由22.0FPS提高到40.6FPS，能够满足移动场景下GPR地下目标实时探测任务的需求。
- 探地雷达 /
- 深度学习 /
- 雷达图像处理 /
- 模型部署 /
- 实时检测
Abstract: Ground Penetrating Radar (GPR) image detection currently faces challenges such as low accuracy, false detections, and missed detections. To overcome these challenges, we propose a novel model referred to as GDS-YOLOv8n for detecting common underground targets in GPR images. The model incorporates the DRRB (Dilated Residual Reparam Block) feature extraction module to achieve enhanced multiscale feature extraction, with certain C2f modules in the YOLOv8n architecture being effectively replaced. In addition, the space-to-depth Conv downsampling module is used to replace the Conv modules corresponding to feature maps with a resolution of 320×320 pixels and less. This replacement assists in mitigating information loss during the downsampling of GPR images, particularly for images with limited resolution and small targets. Furthermore, the detection performance is enhanced using an auxiliary training module, ensuring performance improvement without increasing inference complexity. The introduction of the Inner-SIoU loss function refines bounding box predictions by imposing new constraints tailored to GPR image characteristics. Experimental results on real-world GPR datasets demonstrate the effectiveness of the GDS-YOLOv8n model. For six classes of common underground targets, including metal pipes, PVC pipes, and cables, the model achieves a precision of 97.1%, recall of 96.2%, and mean average precision at 50% IoU (mAP50) of 96.9%. These results indicate improvements of 4.0%, 6.1%, and 4.1%, respectively, compared to corresponding values of the YOLOv8n model, with notable improvements observed when detecting PVC pipes and cables. Compared with those of models such as YOLOv5n, YOLOv7-tiny, and SSD (single shot multibox detector), our model’s mAP50 is improved by 7.20%, 5.70%, and 14.48%, respectively. Finally, the application of our model on a NVIDIA Jetson Orin NX embedded system results in an increase in the detection speed from 22 to 40.6 fps after optimization via TensorRT and FP16 quantization, meeting the demands for the real-time detection of underground targets in mobile scenarios.
- Ground Penetrating Radar (GPR) /
- Deep learning /
- Radar image processing /
- Model deployment /
- Real-time detection

HTML全文

图 1 试验数据采集部分场景

Figure 1. Partial scenarios for experimental data collection

下载: 全尺寸图片幻灯片

图 2 实测B-scan图片

Figure 2. Measured B-scan image

下载: 全尺寸图片幻灯片

图 3 GDS-YOLOv8n网络结构图

Figure 3. GDS-YOLOv8n network structure diagram

下载: 全尺寸图片幻灯片

图 4 DRRB模块

Figure 4. DRRB module

下载: 全尺寸图片幻灯片

图 5 DRB模块

Figure 5. DRB module

下载: 全尺寸图片幻灯片

图 6 结构重参数化过程

Figure 6. Structural reparameterization process

下载: 全尺寸图片幻灯片

图 7 有效感受野

Figure 7. Effective receptive field

下载: 全尺寸图片幻灯片

图 8 SPD-Conv模块

Figure 8. SPD-Conv module

下载: 全尺寸图片幻灯片

图 9 带辅助训练头的训练方式

Figure 9. Training method of auxiliary head

下载: 全尺寸图片幻灯片

图 10 训练过程曲线

Figure 10. Training process curve

下载: 全尺寸图片幻灯片

图 11 GPR B-scan图像检测结果对比

Figure 11. Comparison of GPR B-scan image detection results

下载: 全尺寸图片幻灯片

图 12 测试集上YOLOv8n模型的精确度-召回率曲线

Figure 12. P-R curve of YOLOv8n model on the test dataset

下载: 全尺寸图片幻灯片

图 13 测试集上GSD-YOLOv8n模型的精确度-召回率曲线

Figure 13. P-R curve of GSD-YOLOv8n model on the test dataset

下载: 全尺寸图片幻灯片

图 14 不同模型在测试集上的精确度-召回率曲线

Figure 14. P-R Curves of Different Models on the Test Set

下载: 全尺寸图片幻灯片

图 15 模型综合性能比较

Figure 15. Comprehensive performance comparison of models

下载: 全尺寸图片幻灯片

图 16 模型部署流程

Figure 16. Model deployment process

下载: 全尺寸图片幻灯片

表 1 DRB模块参数

Table 1. DRB module parameter

size	k	d
5	3, 3	1, 2
7	5, 5, 3	1, 2, 3
9	5, 5, 3, 3	1, 2, 3, 4
11	5, 5, 3, 3, 3	1, 2, 3, 4, 5
13	5, 7, 3, 3, 3	1, 2, 3, 4, 5

下载: 导出CSV

表 2 DRRB模块性能对比

Table 2. DRRB module performance comparison

20×20	40×40	Params (Million)	mAP50 (%)
(5,7)	(7,9)	2.69	93.9
(5,7)	(9,11)	2.699	94
(7,9)	(9,11)	2.706	95.3
(7,9)	(11,13)	2.717	94.8
(9,11)	(11,13)	2.727	94.1
注：(x, y)表示使用尺寸为x和y的DRB模块。

下载: 导出CSV

表 3 模型训练参数设置

Table 3. Model Training Parameter Settings

参数名称	参数设置
迭代次数	300
批大小	16
早停机制	50
初始学习率	0.01
最终学习率	0.0001
动量	0.937
优化器权重衰减	0.0005
优化器	auto
注：优化器参数设置auto时，自动选择AdamW优化器。

下载: 导出CSV

表 4 算法改进前后P, R值对比

Table 4. Comparison of P and R values before and after algorithm improvement

检测目标	性能指标	YOLOv8n	GDS-YOLOv8n
金属管	P(%)	95.5	96.3
金属管	R(%)	82.1	87.2
PVC管	P(%)	93.4	97.5
PVC管	R(%)	83.2	98.3
电缆线	P(%)	100	100
电缆线	R(%)	85.1	100
含水塑料瓶	P(%)	93.2	100
含水塑料瓶	R(%)	95.3	100
注浆不密实加固体空洞	P(%)	97.4	97.7
注浆不密实加固体空洞	R(%)	96.5	93
管网	P(%)	79.5	91.2
管网	R(%)	98.4	98.4
总计	P(%)	93.1	97.1
总计	R(%)	90.1	96.2
注：加粗数字表示最优值。

下载: 导出CSV

表 5 消融实验结果

Table 5. Results of ablation experiments

实验	模型	P (%)	R (%)	mAP50 (%)	mAP50:95 (%)
1	YOLOv8n	93.1	90.1	92.8	60.7
2	YOLOv8n+A	95.1	93.2	95.3	63.4
3	YOLOv8n+B	94.6	92.6	94.4	61.3
4	YOLOv8n+C	96.5	91.3	94.7	61.9
5	YOLOv8n+D(ratio=1.29)	95.2	92	94.3	61.8
6	YOLOv8n+A+B	96.4	94.5	96	64
7	YOLOv8n+A+B+C	96.8	94.4	96.5	64.2
8	YOLOv8n+A+B+C+L(ratio=1.28)	97.1	96.2	96.9	64.3
注：加粗数字表示最优值，ratio是Inner-SIoU中的比例因子。

下载: 导出CSV

表 6 对比实验结果

Table 6. Results of comparison experiments

模型	Params(Million)	GFLOPs	Model Size(MB)	P (%)	R (%)	mAP50 (%)	mAP50:90 (%)	FPS (bs=1)
SSD^*	24.28	176.23	93.1	84.03	81.86	82.42	48.3	35.2
YOLOv5n	1.77	4.2	3.68	89.3	86	89.7	61.1	129.9
YOLOv5s	7.03	15.8	13.7	90.7	90.4	92.3	62.8	123.5
YOLOv7-tiny	6.03	13.1	11.7	91.9	92.6	91.2	57.6	80.0
YOLOv8n	3.01	8.1	5.94	93.1	90.1	92.8	60.7	109.9
YOLOv8s	11.13	28.4	21.4	94.7	90.7	93.7	62.0	103.1
YOLOv8m	25.8	78.7	49.5	94.1	96.3	96.7	63.3	55.6
GDS-YOLOv8n	4.43	11.2	10.3	97.1	96.2	96.9	64.3	97.2
注：加粗数字表示最优值；bs=1表示 batch size取1时的值；$* $表示输入图片尺寸是512×512。

下载: 导出CSV

表 7 GDS-YOLOv8n模型部署格式性能对比

Table 7. Performance comparison of GDS-YOLOv8n model deployment formats

模型	Model Size (MB)	mAP50 (%)	FPS (bs=1)
GDS-YOLOv8n.pt	10.3	96.9	22
GDS-YOLOv8n.onnx	17.1	96.9	6.5
GDS-YOLOv8n.engine(FP32)	19.1	96.9	31.4
GDS-YOLOv8n.engine(FP16)	10.7	96.9	40.6
GDS-YOLOv8n.engine(INT8)	5.6	84.1	46
注：FP32, FP16和INT8分别表示以32位、16位浮点数和8位整数量化模型。

下载: 导出CSV

参考文献(21)

[1]	LIU Wenchao, LUO Rong, XIAO Manzhe, et al. Intelligent detection of hidden distresses in asphalt pavement based on GPR and deep learning algorithm[J]. Construction and Building Materials, 2024, 416: 135089. doi: 10.1016/j.conbuildmat.2024.135089.
[2]	ZHU Jiasong, ZHAO Dingyi, and LUO Xianghuan. Evaluating the optimised YOLO-based defect detection method for subsurface diagnosis with ground penetrating radar[J]. Road Materials and Pavement Design, 2024, 25(1): 186–203. doi: 10.1080/14680629.2023.2199880.
[3]	RICHARDS E, STUEFER S, RANGEL R C, et al. An evaluation of GPR monitoring methods on varying river ice conditions: A case study in Alaska[J]. Cold Regions Science and Technology, 2023, 210: 103819. doi: 10.1016/j.coldregions.2023.103819.
[4]	WANG Xiaofen, WANG Peng, ZHANG Xiaotong, et al. Target electromagnetic detection method in underground environment: A review[J]. IEEE Sensors Journal, 2022, 22(14): 13835–13852. doi: 10.1109/JSEN.2022.3175502.
[5]	KAUR P, DANA K J, ROMERO F A, et al. Automated GPR rebar analysis for robotic bridge deck evaluation[J]. IEEE Transactions on Cybernetics, 2016, 46(10): 2265–2276. doi: 10.1109/TCYB.2015.2474747.
[6]	SARKAR M, SR N, HEMANI M, et al. Parameter efficient local implicit image function network for face segmentation[C]. The IEEE/CVF Conference on Computer Vision and Pattern Recognition, Vancouver, Canada, 2023: 20970–20980. doi: 10.1109/CVPR52729.2023.02009.
[7]	TONG Zheng, GAO Jie, and YUAN Dongdong. Advances of deep learning applications in ground-penetrating radar: A survey[J]. Construction and Building Materials, 2020, 258: 120371. doi: 10.1016/j.conbuildmat.2020.120371.
[8]	REN Shaoqing, HE Kaiming, GIRSHICK R, et al. Faster R-CNN: Towards real-time object detection with region proposal networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6): 1137–1149. doi: 10.1109/TPAMI.2016.2577031.
[9]	LIU Wei, ANGUELOV D, ERHAN D, et al. SSD: Single shot multibox detector[C]. The 14th European Conference on Computer Vision, Amsterdam, The Netherlands, 2016: 21–37. doi: 10.1007/978-3-319-46448-0_2.
[10]	REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: Unified, real-time object detection[C]. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, USA, 2016: 779–788. doi: 10.1109/CVPR.2016.91.
[11]	杨必胜, 宗泽亮, 陈驰, 等. 车载探地雷达地下目标实时探测法[J]. 测绘学报, 2020, 49(7): 874–882. doi: 10.11947/j.AGCS.2020.20190293. YANG Bisheng, ZONG Zeliang, CHEN Chi, et al. Real time approach for underground objects detection from vehicle-borne ground penetrating radar[J]. Acta Geodaetica et Cartographica Sinica, 2020, 49(7): 874–882. doi: 10.11947/j.AGCS.2020.20190293.
[12]	XIONG Xuetang and TAN Yiqiu. Deep learning-based detection of tie bars in concrete pavement using ground penetrating radar[J]. International Journal of Pavement Engineering, 2023, 24(2): 2155648. doi: 10.1080/10298436.2022.2155648.
[13]	HU Haobang, FANG Hongyuan, WANG Niannian, et al. Defects identification and location of underground space for ground penetrating radar based on deep learning[J]. Tunnelling and Underground Space Technology, 2023, 140: 105278. doi: 10.1016/j.tust.2023.105278.
[14]	王惠琴, 罗佳, 何永强, 等. 改进YOLOv5的探地雷达常见地下管线识别[J]. 地球物理学报, 2024, 67(9): 3588–3604. doi: 10.6038/cjg2023R0431. WANG Huiqin, LUO Jia, HE Yongqiang, et al. Identification of common underground pipelines by ground penetrating radar based on improved YOLOv5[J]. Chinese Journal of Geophysics, 2024, 67(9): 3588–3604. doi: 10.6038/cjg2023R0431.
[15]	刘震, 顾兴宇, 李骏, 等. 探地雷达数值模拟与道路裂缝图像检测的深度学习增强方法[J]. 地球物理学报, 2024, 67(6): 2455–2471. doi: 10.6038/cjg2023R0090. LIU Zhen, GU Xingyu, LI Jun, et al. Deep learning-enhanced numerical simulation of ground penetrating radar and image detection of road cracks[J]. Chinese Journal of Geophysics, 2024, 67(6): 2455–2471. doi: 10.6038/cjg2023R0090.
[16]	DING Xiaohan, ZHANG Xiangyu, HAN Jungong, et al. Scaling up your kernels to 31×31: Revisiting large kernel design in CNNs[C]. The IEEE/CVF Conference on Computer Vision and Pattern Recognition, New Orleans, USA, 2022: 11953–11965. doi: 10.1109/CVPR52688.2022.01166.
[17]	WEI Haoran, LIU Xu, Xu Shouchun, et al. DWRSeg: Rethinking efficient acquisition of multi-scale contextual information for real-time semantic segmentation[J]. arXiv: 2212.01173, 2022. doi: 10.48550/arXiv.2212.01173.
[18]	DING Xiaohan, ZHANG Yiyuan, GE Yixiao, et al. UniRepLKNet: A universal perception large-kernel convnet for audio, video, point cloud, time-series and image recognition[C]. The IEEE/CVF Conference on Computer Vision and Pattern Recognition, Seattle, USA, 2024: 5513–5524. doi: 10.1109/CVPR52733.2024.00527.
[19]	SUNKARA R and LUO Tie. No more strided convolutions or pooling: A new CNN building block for low-resolution images and small objects[C]. Joint European Conference on Machine Learning and Knowledge Discovery in Databases, Grenoble, France, 2023: 443–459. doi: 10.1007/978-3-031-26409-2_27.
[20]	WANG C Y, BOCHKOVSKIY A, and LIAO H Y M. YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors[C]. The IEEE/CVF Conference on Computer Vision and Pattern Recognition, Vancouver, Canada, 2023: 7464–7475. doi: 10.1109/CVPR52729.2023.00721.
[21]	ZHANG Hao, XU Cong, and ZHANG Shuaijie. Inner-IoU: More effective intersection over union loss with auxiliary bounding box[J]. arXiv: 2311.02877, 2023. doi: 10.48550/arXiv.2311.02877.