Concrete bridge crack detection method based on improved YOLO v8s in complex backgrounds
-
摘要: 针对混凝土桥梁裂缝因背景复杂、细小模糊导致检测精度较差的问题，提出了一种基于改进YOLO v8s的混凝土桥梁裂缝检测算法, 精确定位了裂缝位置信息，并测量了裂缝长、宽等关键参数；以YOLO v8s模型为基础，引入全维度动态卷积(ODConv)获取特征图中更丰富的上下文信息，以增强目标特征提取能力，提高对细小模糊裂缝的检测能力; 采用改进通道注意力模块构建了级联双层特征改进注意力(C2f-MA)融合模块，以挖掘特征图中更多的纹理信息，进一步使网络更加关注裂缝特征，从而抑制无关背景信息的干扰，提高复杂背景下的裂缝检测效果；使用焦距交并比(WIoU)损失函数来解决低质量样本的识别问题，进一步优化了模型的收敛速度和检测准确率；在桥梁检测报告中筛选出存在裂缝细小模糊、阴影、人工画线、杂草等具有复杂背景的裂缝图像，通过人工标注的方式建立了复杂背景条件下桥梁裂缝图像数据集；以召回率、平均精度和模型存储容量作为量化评价指标，并依次通过对比试验及消融试验来对模型进行综合评估。研究结果表明：改进YOLO v8s算法的召回率、平均精度和模型存储容量分别为0.829、0.893和11.14 MB，其综合评价指标优于基准方法YOLO v8s和其他目标检测模型，证明了提出的算法在复杂背景下具有良好的鲁棒性。
-
关键词:
- 桥梁工程 /
- 改进YOLO v8s算法 /
- 深度学习 /
- 裂缝检测 /
- 目标检测
Abstract: To address the issue of low detection accuracy for cracks in concrete bridges caused by complex backgrounds, as well as small and obscure features, the crack location information was accurately located, and key parameters such as crack length and width were measured based on the improved YOLO v8s algorithm for concrete bridge crack detection. Based on the YOLO v8s model, the omni-dimensional dynamic convolution (ODConv) was incorporated to capture richer contextual information in feature maps, enhancing the model's ability to extract the target features and detect small and obscure cracks. An improved channel attention module was used to develop the concatenated two-layer feature-modified attention (C2f-MA) fusion module, enabling the extraction of more texture information from feature maps. This modification further made the network focus on crack features, suppressing interferences from irrelevant background information and improving the crack detection performance in complex backgrounds. The weighted intersection over union (WIoU) loss function was introduced to address the challenge of low-quality sample recognition, optimizing the model's convergence speed and detection accuracy. Crack images with complex backgrounds such as small and obscure cracks, shadows, artificial lines, and weeds were screened in the bridge detection report. A bridge crack image dataset was established by manual annotation with complex background conditions. The model's performance was comprehensively evaluated through comparative and ablation experiments by taking recall, average precision and model storage capacity as quantitative evaluation indicators. Research results demonstrate that the improved YOLO v8s algorithm achieves recall, average precision of 0.829, 0.893 and 0.631, respectively, as well as the model storage capacity of 11.14 MB. Its comprehensive evaluation indicators outperforms the baseline YOLO v8s and other target detection models, validating that the proposed algorithm exhibits robust performance in complex backgrounds.-
Key words:
- bridge engineering /
- improved YOLO v8s algorithm /
- deep learning /
- crack detection /
- target detection
-
图 8 多种目标检测算法对比结果
Figure 8. Comparison of results from multiple target detection algorithms
图 11 沿主骨架线方向的裂缝宽度热力图
Figure 11. Crack width heatmap along main skeleton line direction
表 1 试验环境
Table 1. Experimental environment
名称 型号 显卡(GPU) NVIDIA Quadro RTX 3090 显存/GB 24 操作系统 Ubuntu 20.04 Pytorch 1.10.0 Python 3.8.16 CUDA 11.3 Torchvision 0.11.1 
下载: 导出CSV
表 2 YOLO v8s与改进YOLO v8s性能对比
Table 2. Performance comparison between YOLO v8s and improved YOLO v8s
方法 R P0.5 P0.5∶0.95 帧率/(帧·s-1) 模型存储容量/MB YOLO v8s 0.789 0.857 0.586 139 11.13 改进YOLO v8s 0.829 0.893 0.631 123 11.14
下载: 导出CSV
表 3 消融试验结果
Table 3. Results of ablation experiments
试验编号 WIoU ODConv C2f-MA R P0.5 P0.5∶0.95 帧率/(帧·s-1) 模型存储容量/MB 1 × × × 0.789 0.857 0.586 139 11.13 2 √ × × 0.821 0.871 0.595 143 11.13 3 × √ × 0.823 0.872 0.611 135 11.14 4 × × √ 0.808 0.878 0.600 128 11.13 5 √ √ × 0.827 0.879 0.613 141 11.14 6 √ √ √ 0.829 0.893 0.631 123 11.14
下载: 导出CSV
表 4 C2f-MA模块不同位置消融对比结果
Table 4. Comparison of ablation experiment results of C2f-MA module at different locations
试验编号 R P0.5 P0.5∶0.95 帧率/(帧·s-1) 模型存储容量/MB 1 0.823 0.884 0.619 130 11.13 2 0.823 0.887 0.623 132 11.13 3 0.829 0.893 0.631 123 11.14
下载: 导出CSV
表 5 WIoU超参数对试验结果的影响
Table 5. Effects of WIoU hyperparameters on experimental results
α δ R P0.5 P0.5∶0.95 2.3 2 0.833 0.887 0.620 1.9 3 0.810 0.877 0.605 1.6 4 0.829 0.893 0.631 1.4 5 0.825 0.890 0.627 1.3 6 0.830 0.886 0.625
下载: 导出CSV
表 6 多种目标检测算法结果对比
Table 6. Comparison of results from multiple target detection algorithms
方法 R P0.5 P0.5∶0.95 模型存储容量/MB Faster R-CNN 0.302 0.449 0.232 41.35 Deformable DETR 0.582 0.810 0.464 40.10 YOLO v8-GAM-Wise-IoU 0.763 0.832 0.544 12.87 DINO 0.687 0.858 0.585 47.54 改进YOLO v8s 0.829 0.893 0.631 11.14
下载: 导出CSV
-
[1] 陈浩瀚. 混凝土路面图像的裂缝检测算法研究[D]. 东莞: 东莞理工学院, 2023.CHEN Hao-han. Research on crack detection algorithm of concrete pavement images[D]. Dongguan: Dongguan University of Technology, 2023. (in Chinese) [2] 邓露, 褚鸿鹄, 龙砺芝, 等. 基于深度学习的土木基础设施裂缝检测综述[J]. 中国公路学报, 2023, 36(2): 1-21. doi: 10.3969/j.issn.1001-7372.2023.02.001DENG Lu, CHU Hong-hu, LONG Li-zhi, et al. Review of deep learning-based crack detection for civil infrastructure[J]. China Journal of Highways and Transport, 2023, 36(2): 1-21. (in Chinese) doi: 10.3969/j.issn.1001-7372.2023.02.001 [3] 《中国公路学报》编辑部. 中国桥梁工程学术研究综述·2021[J]. 中国公路学报, 2021, 34(2): 1-97. doi: 10.3969/j.issn.1001-7372.2021.02.002Editorial Department of China Journal of Highway and Transport. Review on China's bridge engineering research: 2021[J]. China Journal of Highway and Transport, 2021, 34(2): 1-97. (in Chinese) doi: 10.3969/j.issn.1001-7372.2021.02.002 [4] HSIEH Y A, TSAI Y J. Machine learning for crack detection: review and model performance comparison[J]. Journal of Computing in Civil Engineering, 2020, 34(5): 04020038. doi: 10.1061/(ASCE)CP.1943-5487.0000918 [5] 刘宇飞, 樊健生, 聂建国, 等. 结构表面裂缝数字图像法识别研究综述与前景展望[J]. 土木工程学报, 2021, 54(6): 79-98.LIU Yu-fei, FAN Jian-sheng, NIE Jian-guo, et al. Review and prospect of digital-image-based crack detection of structure surface[J]. China Civil Engineering Journal, 2021, 54(6): 79-98. (in Chinese) [6] MUNAWAR H S, HAMMAD A W A, HADDAD A, et al. Image-based crack detection methods: a review[J]. Infrastructures, 2021, 6(8): 115. doi: 10.3390/infrastructures6080115 [7] LECUN Y L, BENGIO Y, HINTON G. Deep learning[J]. Nature, 2015, 521(7553): 436-444. doi: 10.1038/nature14539 [8] 沙爱民, 童峥, 高杰. 基于卷积神经网络的路表病害识别与测量[J]. 中国公路学报, 2018, 31(1): 1-10. doi: 10.3969/j.issn.1001-7372.2018.01.001SHA Ai-min, TONG Zheng, GAO Jie. Recognition and measurement of pavement diseases based on convolutional neural networks[J]. China Journal of Highway and Transport, 2018, 31(1): 1-10. (in Chinese) doi: 10.3969/j.issn.1001-7372.2018.01.001 [9] DENG Li, CHU Han-han, SHI Peng, et al. Region-based CNN method with deformable modules for visually classifying concrete cracks[J]. Applied Sciences, 2020, 10(7): 2528. doi: 10.3390/app10072528 [10] 晏班夫, 徐观亚, 栾健, 等. 基于Faster R-CNN与形态法的路面病害识别[J]. 中国公路学报, 2021, 34(9): 181-193. doi: 10.3969/j.issn.1001-7372.2021.09.015YAN Ban-fu, XU Guan-ya, LUAN Jian, et al. Pavement distress detection based on faster R-CNN and morphological operations[J]. China Journal of Highway and Transport, 2021, 34(9): 181-193. (in Chinese) doi: 10.3969/j.issn.1001-7372.2021.09.015 [11] 余加勇, 李锋, 薛现凯, 等. 基于无人机及Mask R-CNN的桥梁结构裂缝智能识别[J]. 中国公路学报, 2021, 34(12): 80-90. doi: 10.3969/j.issn.1001-7372.2021.12.007YU Jia-yong, LI Feng, XUE Xian-kai, et al. Intelligent identification of bridge structure cracks based on unmanned aerial vehicle and mask R-CNN[J]. China Journal of Highway and Transport, 2021, 34(12): 80-90. (in Chinese) doi: 10.3969/j.issn.1001-7372.2021.12.007 [12] REN Shao-qing, HE Kai-ming, 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 (TPAMI), 2017, 39(6): 1137-1149. doi: 10.1109/TPAMI.2016.2577031 [13] LIU Wei, ANGUELOV D, ERHAN D, et al. SSD: single shot multi-box detector[C]//Springer. Proceedings of the 2016 European Conference on Computer Vision. Berlin: Springer, 2016: 21-37. [14] REDMON J, FARHADI A. YOLOv3: an incremental improvement[J]. 2018, DOI: 10.48550/arXiv.804.02767 .[15] 彭雨诺, 刘敏, 万智, 等. 基于改进YOLO的双网络桥梁表观病害快速检测算法[J]. 自动化学报, 2022, 48(4): 1018-1032.PENG Yu-nuo, LIU Min, WAN Zhi, et al. A dual deep network based on the improved YOLO for fast bridge surface defect detection[J]. Acta Automatica Sinica, 2022, 48(4): 1018-1032. (in Chinese) [16] HUSSAIN M. YOLO-v1 to YOLO-v8, the rise of YOLO and its complementary nature toward digital manufacturing and industrial defect detection[J]. Machines, 2023, 11(7): 677. doi: 10.3390/machines11070677 [17] SWATHI Y, CHALLA M. YOLO v8: advancements and innovations in object detection[C]//Springer. International Conference on Smart Computing and Communication. Berlin: Springer, 2024: 1-13. [18] BOCHKOVSKIY A, WANG C Y, LIAO H Y M. YOLOv4: optimal speed and accuracy of object detection[J]. 2020, DOI: 10.48850/arXiv.2004.10934 .[19] ZHENG Zhao-hui, WANG Ping, REN Dong-wei, et al. Enhancing geometric factors in model learning and inference for object detection and instance segmentation[J]. IEEE Transactions on Cybernetics, 2022, 52(8): 8574-8586. doi: 10.1109/TCYB.2021.3095305 [20] ZHENG Zhao-hui, WANG Peng, LIU Wei, et al. Distance-IoU loss: faster and better learning for bounding box regression[C]//AAAI. Proceedings of the AAAI Conference on Artificial Intelligence. New York: AAAI, 2020: 12993-13000. [21] TONG Zan-jia, CHEN Yu-hang, XU Ze-wei, et al. Wise-IoU: bounding box regression loss with dynamic focusing mechanism[J]. 2023, DOI: 10.48550/arXiv.2301.10051 .[22] LI Chao, ZHOU Ao-jun, YAO An-bang. Omni-dimensional dynamic convolution[J]. 2022, arXiv preprint arXiv: 2209.07947. [23] WANG Qi-long, WU Bang-gu, ZHU Peng-fei, et al. ECA-Net: efficient channel attention for deep convolutional neural networks[C]//IEEE. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2020: 11534-11542. [24] HU Jie, SHEN Li, ALBANIE S, et al. Squeeze-and- excitation networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020, 42(8): 2011-2023. doi: 10.1109/TPAMI.2019.2913372 [25] HE Tong, ZHANG Zhi, ZHANG Hang, et al. Bag of tricks for image classification with convolutional neural networks[C]//IEEE. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2019: 558-567. [26] ZHU Xi-zhou, SU Wei-jie, LU Le-wei, et al. Deformable DETR: deformable transformers for end-to-end object detection[J]. 2020, DOI: 10.48550/arXiv.2010.04159 .[27] ZHANG Hao, LI Feng, LIU Shi-long, et al. DINO: DETR with improved denoising anchor boxes for end-to-end object detection[J]. 2022, DOI: 10.48550/arXiv.2203.03605 .[28] XONG Chen-qi, ZATED T, ABDELKADER E M. A novel YOLO v8-GAM-Wise-IoU model for automated detection of bridge surface cracks[J]. Construction and Building Materials, 2024, 414: 135025. doi: 10.1016/j.conbuildmat.2024.135025 [29] HANG Die, YANG Jian-xi, JIANG Shi-xin, et al. Lightweight mesh crack detection algorithm based on efficient attention mechanism[J]. International Journal of Robotics and Automation, 2023, 38(3): 170-179. [30] ZHANG T Y, SUEN C Y. A fast parallel algorithm for thinning digital patterns[J]. Communications of the ACM, 1984, 27(3): 236-239. doi: 10.1145/357994.358023 -
点击查看大图
图(12) / 表(6)
计量
- 文章访问数: 76
- HTML全文浏览量: 14
- PDF下载量: 7
- 被引次数: 0
