Intelligent detection method for minor defects of permafrost roads based on low-altitude UAV inspection
-
摘要:
针对冻土区公路表面细微病害在无人机巡检中尺度小、对比度低、背景复杂等问题,提出了一种轻量化智能检测方法(YOLO-ESG)。主干网络采用EfficientNet v2,通过MBConv模块与SE注意力机制,在控制参数量的同时增强特征表达能力;颈部网络引入Slim-neck结构,融合VoVGSCSP模块与GSConv卷积,构建了高效的多尺度特征融合路径,以提升小目标特征提取与融合能力。结果表明:提出的智能检测方法在测试集上mAP50达到0.917,精确率、召回率与F1分数分别为0.891、0.885与0.888;与多种主流轻量化主干网络相比,EfficientNet v2在精度与效率平衡上表现最优;颈部网络消融试验表明,Slim-neck结合GSConv与VoVGSCSP模块能进一步提升模型对细微病害的检测性能。本方法在保持较高推理速度(18.4 ms·帧-1)的同时,实现了对冻土公路表面细微病害的高精度、强鲁棒检测,可为无人机自动化巡检、冻土路基融沉早期预警及智能养护提供支撑。
Abstract:To address the problems of small scale, low contrast, and complex background of minor defects on surface of permafrost roads during unmanned aerial vehicle (UAV) inspection, a lightweight intelligent detection method was proposed. EfficientNet v2 was adopted as the backbone network. It enhanced the feature representation capability while controlling the parameter quantity through mobile inverted bottleneck convolution modules and the squeeze-and-excitation attention mechanism. The Slim-neck structure was introduced into the neck network. The structure was combined with the variety of view group shuffle cross stage partial network (VoVGSCSP) module and grouped spatial convolution (GSConv) to construct an efficient multi-scale feature fusion path, improving the feature extraction and fusion capability for small targets. Results show that the proposed intelligent detection method achieves mAP50 of 0.917 on the test set, with precision, recall, and F1-score of 0.891, 0.885, and 0.888, respectively. Compared with various mainstream lightweight backbone networks, EfficientNet v2 achieves the best balance between accuracy and efficiency. The ablation studies on the neck network indicate that combining Slim-neck with GSConv and VoVGSCSP further improves model detection performance for minor defects. The proposed method achieves high-accuracy and highly robust detection of minor defects on surface of permafrost roads while maintaining a high inference speed of 18.4 ms·frame-1, providing support for UAV automated inspection, early warning of permafrost roadbed thaw settlement, and intelligent maintenance.
-
Key words:
- road engineering /
- permafrost road /
- deep learning /
- minor defect /
- low-altitude detection /
- UAV inspection
-
表 1 网络配置
Table 1. Network configuration
层号 模块 n s C1 C2 层数 SE 0 Conv 3×3 1 3 24 1 1 Fused-MBConv 1 1 24 24 2 否 2 Fused-MBConv 4 2 24 48 4 否 3 Fused-MBConv 4 2 48 64 4 否 4 MBConv 4 2 64 128 6 是 5 MBConv 6 1 128 160 9 是 6 MBConv 6 2 160 256 15 是 表 2 冻土区公路病害数据集数据量
Table 2. Data volume of permafrost roads disease dataset
类别 纵向裂缝 横向裂缝 龟裂 坑洞 修复 总计 图像/张 1 729 1 332 486 165 226 2 606 病害/处 3 496 2 810 641 236 283 7 466 表 3 测试结果
Table 3. Test results
参数 图像数量/张 病害数量/处 P R F1 mAP50 纵向裂缝 351 726 0.824 0.800 0.812 0.854 横向裂缝 266 589 0.809 0.774 0.791 0.813 龟裂 109 140 0.904 0.943 0.923 0.965 坑洞 31 44 0.948 0.932 0.940 0.966 修复 39 45 0.970 0.978 0.974 0.987 总计 521 1 544 0.891 0.885 0.888 0.917 表 4 不同模型对比
Table 4. Comparison of different models
主干网络 颈部网络 P R F1 mAP50 参数量/106 计算量/GFLOPs 推理速度/(ms·帧-1) 原网络 原网络 0.856 0.878 0.867 0.899 20.1 67.1 15.4 ShuffleNet v1 原网络 0.870 0.796 0.831 0.874 8.7 13.8 14.6 ShuffleNet v2 原网络 0.843 0.830 0.836 0.871 1.9 5.3 13.2 MobileNet v1 原网络 0.830 0.831 0.831 0.875 4.7 13.1 10.5 MobileNet v2 原网络 0.861 0.854 0.857 0.898 3.0 8.0 11.6 MobileNet v3 原网络 0.853 0.836 0.844 0.883 4.4 8.3 12.3 MobileNet v4 原网络 0.866 0.836 0.851 0.886 32.6 38.8 14.7 GhostNet v1 原网络 0.834 0.792 0.812 0.849 4.1 7.0 17.3 GhostNet v2 原网络 0.859 0.785 0.820 0.866 5.0 6.3 19.9 GhostNet v3 原网络 0.862 0.881 0.871 0.906 18.1 20.0 27.5 EfficientNet v1 原网络 0.856 0.832 0.844 0.888 6.5 3.5 13.9 EfficientNet v2 原网络 0.900 0.844 0.871 0.907 21.0 53.0 18.5 EfficientNet v2 Slim-neck 0.872 0.892 0.882 0.914 21.5 54.3 19.5 EfficientNet v2 Slim-neck+GSConv 0.891 0.885 0.888 0.917 21.4 54.1 18.4 表 5 与主流轻量化路面病害检测模型的对比
Table 5. Comparison with mainstream lightweight pavement defect detection models
模型 P R F1 mAP50 参数量/106 计算量/GFLOPs 推理速度/(ms·帧-1) YOLOv5s-GSIM[27] 0.865 0.798 0.802 0.841 4.1 8.8 6.9 LHA-Net[28] 0.870 0.834 0.852 0.884 2.5 5.6 14.9 YOLOv7-RDD[29] 0.850 0.895 6.0 2.0 6.9 YOLOv8n-SBP[30] 0.581 0.573 1.2 3.6 9.1 YOLO-RAPD[31] 0.932 0.768 0.842 0.754 115.9 25.0 本模型 0.891 0.885 0.888 0.917 21.4 54.1 18.4 -
[1] 刘德平, 汪双杰, 金龙. 共和至玉树公路多年冻土地温拟合模型[J]. 中国公路学报, 2015, 28(12): 100-105.LIU De-ping, WANG Shuang-jie, JIN Long. Permafrost ground temperature fitting model of Gonghe-Yushu highway[J]. China Journal of Highway and Transport, 2015, 28(12): 100-105. [2] 温智, 盛煜, 马巍, 等. 退化性多年冻土地区公路路基地温和变形规律[J]. 岩石力学与工程学报, 2009, 28(7): 1477-1483.WEN Zhi, SHENG Yu, MA Wei, et al. Ground temperature and deformation laws of highway embankments in degenerative permafrost regions[J]. Chinese Journal of Rock Mechanics and Engineering, 2009, 28(7): 1477-1483. [3] 李金平, 王佐, 张娟, 等. 多年冻土路基不均匀变形成因分析[J]. 中国公路学报, 2015, 28(12): 78-85, 91.LI Jin-ping, WANG Zuo, ZHANG Juan, et al. Cause analysis on uneven deformation of embankment in permafrost regions[J]. China Journal of Highway and Transport, 2015, 28(12): 78-85, 91. [4] 李金平, 张娟, 陈建兵, 等. 高寒冻土区路基变形演化规律与破坏特征[J]. 交通运输工程学报, 2016, 16(4): 78-87. doi: 10.19818/j.cnki.1671-1637.2016.04.008 LI Jin-ping, ZHANG Juan, CHEN Jian-bing, et al. Evolution laws and failure characteristics of subgrade deformation in alpine permafrost region[J]. Journal of Traffic and Transportation Engineering, 2016, 16(4): 78-87. doi: 10.19818/j.cnki.1671-1637.2016.04.008 [5] 马建, 赵祥模, 贺拴海, 等. 路面检测技术综述[J]. 交通运输工程学报, 2017, 17(5): 121-137.MA Jian, ZHAO Xiang-mo, HE Shuan-hai, et al. Review of pavement detection technology[J]. Journal of Traffic and Transportation Engineering, 2017, 17(5): 121-137. [6] 路鑫, 钱学东. 多年冻土公路沥青路面技术路况检测与评价[J]. 公路, 2024, 69(4): 68-73.LU Xin, QIAN Xue-dong. Inspection and evaluation of technical conditions of asphalt pavement on permafrost highways[J]. Highway, 2024, 69(4): 68-73. [7] CAI Guo-jun, TIAN Hong-liang, LIU Lu-lu, et al. Research progress of intelligent testing technology and evaluation methods for subgrade engineering[J]. Journal of Road Engineering, 2025, 5(2): 164-183. doi: 10.1016/j.jreng.2025.03.001 [8] YANG Xu, ZHANG Jian-qi, LIU Wen-bo, et al. Automation in road distress detection, diagnosis and treatment[J]. Journal of Road Engineering, 2024, 4(1): 1-26. doi: 10.1016/j.jreng.2024.01.005 [9] YANG J S, CHEN W, XIAO X. Pavement damage detection with a novel semi-supervised YOLOv8 network[J]. Measurement, 2026, 257: 118860. doi: 10.1016/j.measurement.2025.118860 [10] 汪双杰, 熊丽, 张驰, 等. 多年冻土区公路病害模糊专家预测方法[J]. 交通运输工程学报, 2016, 16(4): 112-121. doi: 10.19818/j.cnki.1671-1637.2016.04.012WANG Shuang-jie, XIONG Li, ZHANG Chi, et al. Fuzzy expert prediction method for highway diseases in permafrost region[J]. Journal of Traffic and Transportation Engineering, 2016, 16(4): 112-121. doi: 10.19818/j.cnki.1671-1637.2016.04.012 [11] CHEN C, SEO H, JUN C H, et al. Pavement crack detection and classification based on fusion feature of LBP and PCA with SVM[J]. International Journal of Pavement Engineering, 2022, 23(9): 3274-3283. doi: 10.1080/10298436.2021.1888092 [12] PENG C, YANG M Q, ZHENG Q H, et al. A triple-thresholds pavement crack detection method leveraging random structured forest[J]. Construction and Building Materials, 2020, 263: 120080. doi: 10.1016/j.conbuildmat.2020.120080 [13] QU Z, CAO C, LIU L, et al. A deeply supervised convolutional neural network for pavement crack detection with multiscale feature fusion[J]. IEEE Transactions on Neural Networks and Learning Systems, 2022, 33(9): 4890-4899. doi: 10.1109/TNNLS.2021.3062070 [14] HAO Z X, ZHU Y H, LU C. Modified YOLOv8 for automated detection and identification towards highway pavement distress[J]. Construction and Building Materials, 2025, 495: 143428. doi: 10.1016/j.conbuildmat.2025.143428 [15] DU Y C, PAN N, XU Z H, et al. Pavement distress detection and classification based on YOLO network[J]. International Journal of Pavement Engineering, 2021, 22(13): 1659-1672. doi: 10.1080/10298436.2020.1714047 [16] ZUO C H, HUANG N X, YUAN C, et al. Pavement-DETR: A high-precision real-time detection transformer for pavement defect detection[J]. Sensors, 2025, 25(8): 2426. doi: 10.3390/s25082426 [17] 翟军治, 孙朝云, 裴莉莉, 等. 多尺度特征增强的路面裂缝检测方法[J]. 交通运输工程学报, 2023, 23(1): 291-308.ZHAI Jun-zhi, SUN Zhao-yun, PEI Li-li, et al. Pavement crack detection method based on multi-scale feature enhancement[J]. Journal of Traffic and Transportation Engineering, 2023, 23(1): 291-308. [18] 高铭, 陈鑫, 蒋烁, 等. 基于自适应微调的露天矿山低空无人机旋转目标检测[J]. 交通运输工程学报, 2026, 26(3): 291-302. doi: 10.19818/j.cnki.1671-1637.2026.158GAO Ming, CHEN Xin, JIANG Shuo, et al. Rotated object detection using low-altitude UAVs for open-pit mines with adaptive fine-tuning[J]. Journal of Traffic and Transportation Engineering, 2026, 26(3): 291-302. doi: 10.19818/j.cnki.1671-1637.2026.158 [19] 马涛, 吴俊, 唐樊龙, 等. 基于多源数据与大模型的无人机巡航风险识别技术[J]. 交通运输工程学报, 2026, 26(3): 75-88. doi: 10.19818/j.cnki.1671-1637.2026.036 MA Tao, WU Jun, TANG Fan-long, et al. Unmanned aerial vehicle cruise risk identification technology based on multi-source data and large models[J]. Journal of Traffic and Transportation Engineering, 2026, 26(3): 75-88. doi: 10.19818/j.cnki.1671-1637.2026.036 [20] 杨炜, 方虹苏, 唐湘松, 等. 桥梁裂缝病害检测的轻量化YOLOv8-ALTE算法[J]. 交通运输工程学报, 2025, 25(6): 75-89. doi: 10.19818/j.cnki.1671-1637.2025.06.007 YANG Wei, FANG Hong-su, TANG Xiang-song, et al. Lightweight YOLOv8-ALTE algorithm for bridge crack disease detection[J]. Journal of Traffic and Transportation Engineering, 2025, 25(6): 75-89. doi: 10.19818/j.cnki.1671-1637.2025.06.007 [21] 周豪, 黄海松. 改进YOLOv8轻量化轴孔检测算法[J]. 组合机床与自动化加工技术, 2025(10): 110-115.ZHOU Hao, HUANG Hai-song. Improved YOLOv8 lightweight shaft hole detection algorithm[J]. Modular Machine Tool & Automatic Manufacturing Technique, 2025 (10): 110-115. [22] ZHUANG X Y, FAN W J, GUO H W, et al. Surrounding rock classification from onsite images with deep transfer learning based on EfficientNet[J]. Frontiers of Structural and Civil Engineering, 2024, 18(9): 1311-1320. doi: 10.1007/s11709-024-1134-7 [23] 赵晋陵, 詹媛媛, 王娟, 等. 基于SE-UNet的冬小麦种植区域提取方法[J]. 农业机械学报, 2022, 53(9): 189-196.ZHAO Jin-ling, ZHAN Yuan-yuan, WANG Juan, et al. SE-UNet-based extraction of winter wheat planting areas[J]. Transactions of the Chinese Society for Agricultural Machinery, 2022, 53(9): 189-196. [24] LI H L, LI J, WEI H B, et al. Slim-neck by GSConv: A lightweight-design for real-time detector architectures[J]. Journal of Real-time Image Processing, 2024, 21(3): 62. doi: 10.1007/s11554-024-01436-6 [25] SONG K, HUANG Q X. GMVG-Net: Lightweight multiscale feature fusion network for welding defect detection[J]. Signal, Image and Video Processing, 2025, 19(6): 436. doi: 10.1007/s11760-024-03738-5 [26] ZHAO X M, SONG Y L. Improved ship detection with YOLOv8 enhanced with MobileViT and GSConv[J]. Electronics, 2023, 12(22): 4666. doi: 10.3390/electronics12224666 [27] WANG H, SUN G S, YANG J P, et al. Efficient pavement damage detection algorithm based on lightweight feature enhancement network[J]. Journal of Performance of Constructed Facilities, 2025, 39(5): 04025049. doi: 10.1061/JPCFEV.CFENG-5120 [28] LI G, ZHANG C, LI M, et al. LHA-Net: A lightweight and high-accuracy network for road surface defect detection[J]. IEEE Transactions on Intelligent Vehicles, 2024, 9(12): 7577-7591. doi: 10.1109/TIV.2024.3400035 [29] NING Z P, WANG H, LI S L, et al. YOLOv7-RDD: A lightweight efficient pavement distress detection model[J]. IEEE Transactions on Intelligent Transportation Systems, 2024, 25(7): 6994-7003. doi: 10.1109/TITS.2023.3347034 [30] YE L, LU S J, HU F K, et al. YOLOv8n-SBP: A lightweight and efficient model for pavement distress detection[J]. Journal of Transportation Engineering, Part B: Pavements, 2025, 151(4): 04025048. doi: 10.1061/JPEODX.PVENG-1815 [31] WEI Y F, PENG Y, CHENG H X, et al. YOLO-RAPD: Enhanced YOLOv8s-based automated detection of road assets and pavement distress[J]. Journal of Computing in Civil Engineering, 2025, 39(6): 04025092. doi: 10.1061/JCCEE5.CPENG-6419 -
下载: