基于YOLOv8-OSRA的钢拱塔表观病害多目标检测方法

doi:10.11996/JG.j.2095-302X.2025061327

摘要/Abstract

摘要： 钢拱塔是钢拱斜拉桥的主要承重结构，其表观病害(如腐蚀、剥落和裂缝等)的早期检测与评估对保障桥梁结构安全至关重要。针对传统人工检测方法效率低下、主观性强且难以覆盖高空隐蔽区域的问题。提出了一种基于改进YOLOv8n-Seg深度学习框架与OSRA注意力机制的智能检测方法。利用自主研发的轮轨式检测机器人系统采集钢拱塔内部高分辨率图像数据，结合开源数据集构建了包含5 846张原始图像的多病害数据集，并通过随机裁剪、镜像翻转和亮度调整等数据增强技术将样本量扩展至23 378张图像。在算法设计层面，创新性地将OSRA注意力模块嵌入YOLOv8n-Seg网络的特征融合层，通过重叠分块策略和局部细化机制显著提升了模型对不规则边界和微小病害特征的捕捉能力。结果表明：优化后的YOLOv8-OSRA模型在独立测试集上取得了显著的性能提升，锈蚀检测mAP@0.5达到90.9% (提升2.6%)，裂缝识别精度达到87.0% (提升1.1%)，剥落检测准确率达到81.9% (提升2.1%)。消融实验进一步验证了OSRA模块在保持计算效率(仅增加0.8% GFLOPs)的同时，其性能优势显著优于SE和CBAM等注意力机制。研究成果为钢拱塔病害检测提供可部署于移动检测设备的轻量化解决方案，且提出的多尺度特征增强方法为复杂钢结构表面缺陷检测提供参考。

关键词: 桥梁工程, 轮轨式检测机器人, 表观病害检测, 机器视觉, 深度学习, YOLOv8

Abstract:

Steel arch towers, as the primary load-bearing structures of steel arch cable-stayed bridges, require early detection and assessment of their surface defects—such as corrosion, spalling, and cracks—to ensure structural safety. However, traditional manual inspection methods are inefficient, highly subjective, and unable to access concealed areas at high altitudes. To address these challenges, an intelligent detection method based on an improved YOLOv8n-Seg deep-learning framework integrated with the OSRA attention mechanism was proposed. High-resolution internal images of steel arch towers were collected using a self-developed rail-guided inspection robot system. A comprehensive dataset containing 5 846 original images was constructed from both collected and open-source data, and data augmentation techniques—including random cropping, mirroring, and brightness adjustment—expanded the dataset to 23 378 images. At the algorithmic level, the OSRA attention module was innovatively embedded into the feature fusion layer of the YOLOv8n-Seg network. By leveraging an overlapping patching strategy and a local refinement mechanism, the model’s ability to capture irregular boundaries and small-scale defect features was significantly enhanced. Experimental results demonstrated that the optimized YOLOv8-OSRA model achieved notable performance improvements on an independent test set: corrosion detection mAP@0.5 reached 90.9% (+2.6%), crack identification accuracy reached 87.0% (+1.1%), and spalling detection accuracy reached 81.9% (+2.1%). Ablation experiments further confirmed that the OSRA module, while maintaining computational efficiency (increasing GFLOPs by only 0.8%), outperformed conventional attention mechanisms such as SE and CBAM. The findings provided a lightweight and deployable solution for steel arch tower defect detection, and the proposed multi-scale feature enhancement approach offered valuable insights for detecting surface defects in complex steel structures.

Key words: bridge engineering, wheel rail inspection robot, surface defect detection, computer vision, deep learning, YOLOv8

中图分类号:

王海涵. 基于YOLOv8-OSRA的钢拱塔表观病害多目标检测方法[J]. 图学学报, 2025, 46(6): 1327-1336.

WANG Haihan. Multi object detection method for surface defects of steel arch towers based on YOLOv8-OSRA[J]. Journal of Graphics, 2025, 46(6): 1327-1336.

图/表 16

图1 C2F模块的架构

Fig. 1 Architecture of the C2F module

图2 YOLOv8-Seg的网络架构

Fig. 2 YOLOv8-Seg network architecture

图3 实桥检测((a) 钢拱塔斜拉桥；(b) 轮轨式检测机器人)

Fig. 3 Real bridge inspection ((a) Steel arch tower cable-stayed bridge; (b) Track-mounted inspection robot)

图4 典型损伤类型((a) 锈蚀；(b) 表观开裂；(c) 涂层脱落)

Fig. 4 Typical damage types ((a) Rust; (b) Apparent cracking; (c) Coating shedding)

图5 数据标注((a) 锈蚀；(b) 裂缝；(c) 涂层脱落)

Fig. 5 Data annotation ((a) Rust; (b) Crack; (c) Coating peeling)

图6 典型损伤图像数据增强((a) 原图；(b) 镜像；(c) 亮度调整；(d) 裁剪)

Fig. 6 Data enhancement of typical damaged images ((a) Original image; (b) Mirror image; (c) Brightness adjustment; (d) Cropping)

图7 YOLOv8-OSRA模型的网络架构

Fig. 7 Network architecture of the YOLOv8-OSRA model

图8 损失值变化曲线((a) 边界框的损失值变化曲线；(b) 分割的损失值变化曲线)

Fig. 8 Loss value change curve ((a) Loss value change curve of bounding box; (b) Loss value change curve of segmentation)

表1 不同网络训练结果对比/%

Table 1 Comparison of training results of different networks/%

类别	边界框		分割掩码
类别	YOLOv8n-Seg	YOLOv8-OSRA	YOLOv8n-Seg	YOLOv8-OSRA
锈蚀	88.3	90.9	77.8	81.5
表观裂缝	85.9	87.0	63.3	65.5
涂层脱落	79.8	81.9	65.6	67.3
mAP@0.5	84.6	86.6	68.9	71.4

图9 YOLOv8n-Seg的PR曲线((a) 边界框；(b) 掩码)

Fig. 9 PR curve of YOLOv8n-Seg ((a) Bounding box; (b) Mask)

图10 YOLOv8-OSRA的PR曲线((a) 边界框；(b) 掩码)

Fig. 10 PR curve of YOLOv8-OSRA ((a) Bounding box; (b) Mask)

表2 计算复杂度对比

Table 2 Comparison of computational complexity

模型	GFLOPs	参数量/M	FPS (RTX3050)
YOLOv8n-Seg	12.0	6.5	58.8
YOLOv8-OSRA	12.1	10.2	51.5

表3 注意力机制消融实验对比

Table 3 Comparison of ablation experiments on attention mechanisms

注意力类别	mAP@0.5 (锈蚀)/%		mAP@0.5 (裂缝)/%		mAP@0.5 (剥落)/%		GFLOPs
注意力类别	边界框	分割掩码	边界框	分割掩码	边界框	分割掩码	GFLOPs
无	88.3	77.8	85.9	63.3	79.8	65.6	12.0
SE	80.1	66.8	81.7	57.7	76.0	61.0	27.6
CBAM	80.9	65.8	82.7	57.6	76.5	62.3	27.7
SRA	78.9	64.6	81.6	57.7	76.0	62.3	27.8
OSRA	90.9	81.5	87.0	65.5	81.9	67.3	12.1

表4 主流网络性能对比

Table 4 Comparison of mainstream networks

模型	mAP@0.5/%		FPS	Params/M	GFLOPs
模型	检测框	分割掩码	FPS	Params/M	GFLOPs
YOLOv5	73.8	58.0	59.5	5.6	11.0
YOLOv7	81.6	66.5	58.8	13.6	47.7
YOLOv8 (基线)	84.6	68.9	58.8	6.5	12.0
YOLOv8-OSRA	86.6	71.4	51.5	10.2	12.1

图11 模型优化前后检测结果对比((a) 原图；(b) 优化前的YOLOv8n-Seg模型检测结果；(c) 优化后的YOLOv8-OSRA模型检测结果)

Fig. 11 Comparison of detection results before and after model optimization ((a) Original image; (b) Detection result of YOLOv8n-Seg model before optimization; (c) Detection result of YOLOv8-OSRA model after optimization)

图12 运算显示窗口((a) 数据展示窗口；(b) 病害信息与定位)

Fig. 12 Operation display window ((a) Data display window; (b) Disease information and location)

参考文献 19

[1]	CHA Y J, CHOI W, SUH G, et al. Autonomous structural visual inspection using region-based deep learning for detecting multiple damage types[J]. Computer-Aided Civil and Infrastructure Engineering, 2018, 33(9): 731-747. DOI URL
[2]	WANG D L, ZHANG Y Q, PAN Y, et al. An automated inspection method for the steel box girder bottom of long-span bridges based on deep learning[J]. IEEE Access, 2020, 8: 94010-94023. DOI URL
[3]	CHEN K H, HUANG Z D, CHEN C, et al. Surface crack detection of steel structures in railroad industry based on multi-model training comparison technique[J]. Processes, 2023, 11(4): 1208. DOI URL
[4]	LUO C, YU L J, YAN J X, et al. Autonomous detection of damage to multiple steel surfaces from 360° panoramas using deep neural networks[J]. Computer-Aided Civil and Infrastructure Engineering, 2021, 36(12): 1585-1599. DOI URL
[5]	LIU R Q, HUANG M, GAO Z M, et al. MSC-DNet: an efficient detector with multi-scale context for defect detection on strip steel surface[J]. Measurement, 2023, 209: 112467. DOI URL
[6]	斯新华, 刘大洋, 张朋, 等. 基于YOLOv4的桥梁钢结构表观病害自动识别方法的研究[J]. 公路, 2022, 67(8): 399-402.
	SI X H, LIU D Y, ZHANG P, et al. Research on the automatic identification method of apparent diseases of bridge steel structure based on YOLOv4[J]. Highway, 2022, 67(8): 399-402 (in Chinese).
[7]	逯鹏, 赵天淞, 王剑, 等. 基于计算机视觉的钢结构表面锈蚀程度检测方法[J]. 工业建筑, 2024, 54(8): 133-139.
	LU P, ZHAO T S, WANG J, et al. A method for detecting surface corrosion degree of steel structures based on computer vision[J]. Industrial Construction, 2024, 54(8): 133-139 (in Chinese).
[8]	王志东, 陈晨阳, 刘晓明. 基于轻量化改进YOLOv8的通信光缆缺陷检测[J]. 图学学报, 2025, 46(1): 28-34. DOI
	WANG Z D, CHEN C Y, LIU X M. The defect detection method for communication optical cables based on lightweight improved YOLOv8[J]. Journal of Graphics, 2025, 46(1): 28-34 (in Chinese). DOI
[9]	崔克彬, 耿佳昌. 基于EE-YOLOv8s的多场景火灾迹象检测算法[J]. 图学学报, 2025, 46(1): 13-27. DOI
	CUI K B, GENG J C. A multi-scene fire sign detection algorithm based on EE-YOLOv8s[J]. Journal of Graphics, 2025, 46(1): 13-27 (in Chinese). DOI
[10]	赵振兵, 韩钰, 唐辰康. 基于改进YOLOv8的配电线路绝缘子缺陷级联检测方法[J]. 图学学报, 2025, 46(1): 1-12. DOI
	ZHAO Z B, HAN Y, TANG C K. Cascade detection method for insulator defects in distribution lines based on improved YOLOv8[J]. Journal of Graphics, 2025, 46(1): 1-12 (in Chinese). DOI
[11]	WANG C Y, BOCHKOVSKIY A, LIAO H Y M. YOLOv7: trainable bag-of-freebies sets new state-of-the-art for real-time object detectors[C]// 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2023: 7464-7475.
[12]	LI C Y, LI L L, JIANG H L, et al. YOLOv6:a single-stage object detection framework for industrial applications[EB/OL]. [2024-05-07]. https://arxiv.org/abs/2209.02976.
[13]	LOU M, ZHANG S, ZHOU H Y, et al. TransXNet: learning both global and local dynamics with a dual dynamic token mixer for visual recognition[EB/OL]. [2024-05-30]. https://arxiv.org/abs/2310.19380.
[14]	WANG W H, XIE E Z, LI X, et al. Pyramid vision transformer: a versatile backbone for dense prediction without convolutions[C]// 2021 IEEE/CVF International Conference on Computer Vision. New York: IEEE Press, 2021: 548-558.
[15]	CHOLLET F. Xception: deep learning with depthwise separable convolutions[C]// 2017 IEEE Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2017: 1800-1807.
[16]	GUO J Y, HAN K, WU H, et al. CMT: convolutional neural networks meet vision transformers[C]// 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2022: 12165-12175.
[17]	LI Y Y, YUAN G, WEN Y, et al. EfficientFormer: vision transformers at MobileNet speed[EB/OL]. [2024-06-25]. https://proceedings.neurips.cc/paper_files/paper/2022/file/5452ad8ee6ea6e7dc41db1cbd31ba0b8-Paper-Conference.pdf.
[18]	HU J, SHEN L, SUN G. Squeeze-and-excitation networks[C]// 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2018: 7132-7141.
[19]	WOO S, PARK J, LEE J Y, et al. CBAM: convolutional block attention module[C]// The 15th European Conference on Computer Vision. Cham: Springer, 2018: 3-19.