VDD-YOLOv9-segm:一种面向车辆损伤的智能检测方法

doi:10.3969/j.issn.1000-1565.2026.03.003

摘要/Abstract

摘要： 针对车辆损伤检测模型中特征提取不足和类别不平衡导致的漏检和误检问题,本文提出了一种面向车辆损伤检测的模型(vehicle damage detection-YOLOv9-segm,VDD-YOLOv9-segm).首先,为有效捕捉特征图间的长程依赖并学习目标类别各特征通道的重要性权重,设计了一种由2个互补的并行子网络组成的双路径结构单元(multi-head squeeze-and-excitation network,MHSENet),并将其嵌入多卷积框架中,以提升模型的特征提取、表达和泛化能力.其次,为了缓解类别间的样本不平衡问题,引入WIoU(Wise Intersection over Union)损失函数替换传统的CIoU损失函数,从而能够有效地衡量模型预测结果和真实标注结果之间的相似度,使模型更好地关注困难样本,提高模型的检测性能.基于自建的10类车辆损伤数据集进行实验,结果表明,所提模型在总体检测和分割效果上均优于其他先进实例分割模型.相较于基线模型,目标框检测精确率提高了3.3个百分点,掩码检测精确率提高了4.3个百分点.消融实验进一步证实了本文模型各模块均对检测效果有显著提升.综合利用MHSENet和WIoU的优势,本文提出的模型有效解决了车辆损伤检测中特征提取不足和类别不平衡导致的漏检、误检问题,显著提升了损伤类别的检测精度.

关键词: 车辆损伤, 实例分割, 注意力机制, 交并比, 骨干网络

Abstract: To address the issues of missed detection and false detection caused by insufficient feature extraction and category imbalance in the vehicle damage model, this paper proposes a model for the vehicle damage detection task(vehicle damage detection-YOLOv9-segm, VDD-YOLOv9-segm). Firstly, in order to effectively capture the long-range dependencies between feature maps and learn the importance weights of - 引用格式:张文恺,杨术明,马永龙,等.基于EDEM的牧场推料机器人参数优化设计与试验［J］.河北大学学报(自然科学版),2026,46(3):225-236.引用格式:翟永杰,宋佳车,陈年昊,等.VDD-YOLOv9-segm:一种面向车辆损伤的智能检测方法［J］.河北大学学报(自然科学版),2026,46(3):249-263.DOI:10.3969/j.issn.1000-1565.2026.03.003VDD-YOLOv9-segm:一种面向车辆损伤的智能检测方法翟永杰¹,宋佳车¹,陈年昊²,王乾铭¹,刘金龙³(1.华北电力大学控制与计算机工程学院,河北保定 071003;2.华南理工大学电力学院,广东广州 510640;3.邦邦汽车销售服务有限公司,北京 100032)摘要:针对车辆损伤检测模型中特征提取不足和类别不平衡导致的漏检和误检问题,本文提出了一种面向车辆损伤检测的模型(vehicle damage detection-YOLOv9-segm,VDD-YOLOv9-segm).首先,为有效捕捉特征图间的长程依赖并学习目标类别各特征通道的重要性权重,设计了一种由2个互补的并行子网络组成的双路径结构单元(multi-head squeeze-and-excitation network,MHSENet),并将其嵌入多卷积框架中,以提升模型的特征提取、表达和泛化能力.其次,为了缓解类别间的样本不平衡问题,引入WIoU(Wise Intersection over Union)损失函数替换传统的CIoU损失函数,从而能够有效地衡量模型预测结果和真实标注结果之间的相似度,使模型更好地关注困难样本,提高模型的检测性能.基于自建的10类车辆损伤数据集进行实验,结果表明,所提模型在总体检测和分割效果上均优于其他先进实例分割模型.相较于基线模型,目标框检测精确率提高了3.3个百分点,掩码检测精确率提高了4.3个百分点.消融实验进一步证实了本文模型各模块均对检测效果有显著提升.综合利用MHSENet和WIoU的优势,本文提出的模型有效解决了车辆损伤检测中特征提取不足和类别不平衡导致的漏检、误检问题,显著提升了损伤类别的检测精度.关键词:车辆损伤;实例分割;注意力机制;交并比;骨干网络中图分类号:TN911.73 文献标志码:A 文章编号:1000-1565(2026)03-0249-15DOI:10.3969/j.issn.1000-1565.2026.03.003 VDD-YOLOv9-segm: An intelligent detection method for vehicle damageZHAI Yongjie¹, SONG Jiache¹, CHEN Nianhao², WANG Qianming¹, LIU Jinlong³(1. School of Control and Computer Engineering, North China Electric Power University, Baoding 071003, China; 2. School of Electric Power, South China University of Technology, Guangzhou 510640, China; 3. Bangbang Automobile Sales and Service Co., Ltd., Beijing 100032, China)Abstract: To address the issues of missed detection and false detection caused by insufficient feature extraction and category imbalance in the vehicle damage model, this paper proposes a model for the vehicle damage detection task(vehicle damage detection-YOLOv9-segm, VDD-YOLOv9-segm). Firstly, in order to effectively capture the long-range dependencies between feature maps and learn the importance weights of - 收稿日期:2025-06-09;修回日期:2025-11-04 基金项目:国家自然科学基金项目(62373151);河北省自然科学基金面上项目(F2023502010);中央高校基本科研业务费专项资金项目(2023JC006;2024MS136) 第一作者:翟永杰(1972—),男,华北电力大学教授,主要从事电力计算机视觉方向研究. E-mail:zhaiyongjie@ncepu.edu.cn 通信作者:陈年昊(1998—),男,华南理工大学在读博士研究生,主要从事电力工程与计算机视觉方向研究. E-mail:815173055@qq.com 第3期翟永杰等:VDD-YOLOv9-segm:一种面向车辆损伤的智能检测方法河北大学学报(自然科学版) 第46卷each feature channel of the target category, a dual-path structure unit(multi-head squeeze-and-excitation network, MHSENet)composed of two complementary parallel subnetworks is designed. And it is embedded in the multi-convolution framework to enhance the feature extraction, expression and generalization capabilities of the model. Secondly, in order to alleviate the sample imbalance problem among categories, the Wise Intersection over Union(WIoU)loss is introduced to replace the original traditional Complete Intersection over Union(CIoU)loss, thereby effectively measuring the similarity between the models predicted results and the real labeled results, enabling the model to better focus on difficult samples and improving the detection performance of the model. The experiment was conducted on a self-built dataset of 10 types of vehicle damage. The comparative experimental results show that the proposed model is superior to other advanced instance segmentation models in both overall detection and segmentation effects. Compared with the baseline model, the precision of the target box has increased by 3.3 percentage points, and the precision of segmentation has increased by 4.3 percentage points. The ablation experiments further confirmed that each module of this paper significantly improved the detection effect. By comprehensively utilizing the advantages of MHSENet and WIoU, the model proposed in this paper effectively solves the problems of missed detection and false detection caused by insufficient feature extraction and category imbalance in vehicle damage detection, and significantly improves the detection accuracy of damage categories.

Key words: vehicle damage, instance segmentation, attention mechanism, Intersection over Union, backbone network

中图分类号:

TN911.73

翟永杰,宋佳车,陈年昊,王乾铭,刘金龙. VDD-YOLOv9-segm:一种面向车辆损伤的智能检测方法[J]. 河北大学学报(自然科学版), 2026, 46(3): 249-263.

ZHAI Yongjie, SONG Jiache, CHEN Nianhao, WANG Qianming, LIU Jinlong. VDD-YOLOv9-segm: An intelligent detection method for vehicle damage[J]. Journal of Hebei University(Natural Science Edition), 2026, 46(3): 249-263.

参考文献

[1] 李子.数据时代新能源汽车保险诈骗案件研究[J].山东警察学院学报, 2023, 35(4): 66-72. DOI:10.3969/j.issn.1673-1565.2023.04.008.
[2] 甘志梅,李根发.保险诈骗罪的立法缺陷及汽车保险欺诈的防治[J].南昌工程学院学报, 2012, 31(5): 54-57. DOI:10.3969/j.issn.1006-4869.2012.05.015.
[3] 谢东升.基于深度学习的车辆智能定损算法研究[D].天津:天津大学, 2019.
[4] Qaddour J, Siddiqa S A. Automatic damaged vehicle estimator using enhanced deep learning algorithm[J]. Intell Syst Appl, 2023, 18: 200192. DOI:10.1016/j.iswa.2023.200192.
[5] 齐玉亮,王伟明,王静,等.基于跨层注意力与多记忆单元的车辆重识别方法[J].河北大学学报(自然科学版), 2025, 45(3): 327-336. DOI:10.3969/j.issn.1000-1565.2025.03.011.
[6] 柯淇晟,花强,张博.基于改进YOLOv8的遥感图像目标检测[J].河北大学学报(自然科学版), 2025, 45(3): 309-316. DOI:10.3969/j.issn.1000-1565.2025.03.009.
[7] 薛林雁,李轩昂,齐晁仪,等.基于改进YOLOv5s的肠镜息肉多分类实时检测方法[J].河北大学学报(自然科学版), 2024, 44(4): 424-432. DOI:10.3969/j.issn.1000-1565.2024.04.011.
[8] Tian X G, Han H. Deep convolutional neural networks with transfer learning for automobile damage image classification[J]. J Database Manag, 2022, 33(3): 1-16. DOI:10.4018/jdm.309738.
[9] Mohammadi Dolat-Abadi H, SarAMI F. An integrated framework for detection of car accident damage using deep neural networks model: A real-life case study[J]. J Ind Syst Eng, 2025, 17(1): 18-35.
[10] Widjojo D, Setyati E, Kristian Y. Integrated deep learning system for car damage detection and classification using deep transfer learning[C] //2022 IEEE 8th Information Technology International Seminar(ITIS). October 19-21, 2022, Surabaya, Indonesia. IEEE, 2022: 21-26. DOI:10.1109/ITIS57155.2022.10010292.
[11] Van Ruitenbeek R E, Bhulai S. Convolutional Neural Networks for vehicle damage detection[J]. Mach Learn Appl, 2022, 9: 100332. DOI:10.1016/j.mlwa.2022.100332.
[12] Peng J B, Dong S B, Yuan H, et al. Car damage detection based on multi-view fusion and alignment: dataset and method[J]. IEEE Trans Intell Transp Syst, 2025, 26(4): 4717-4730. DOI:10.1109/TITS.2025.3542174.
[13] Setyawan H A, Bustamam A, Buyung R A. Analysis performance of one-stage and two stage object detection method for car damage detection[J]. Ijacsa, 2024, 15(7):1064-1073. DOI:10.14569/ijacsa.2024.01507103.
[14] 彭锦波.基于多视角深度学习的车辆损伤检测研究[D].广州:华南理工大学, 2023.
[15] Ren S Q, He K M, Girshick R, et al. Faster R-CNN: towards real-time object detection with region proposal networks[J]. IEEE Trans Pattern Anal Mach Intell, 2017, 39(6): 1137-1149. DOI:10.1109/TPAMI.2016.2577031.
[16] He K, Gkioxari G, DolláR P, et al. Mask r-cnn. 2017: 2961-2969.
[17] Huang Z J, Huang L C, Gong Y C, et al. Mask scoring R-CNN[C] //2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR). June 15-20, 2019, Long Beach, CA, USA. IEEE, 2019: 6402-6411. DOI:10.1109/CVPR.2019.00657.
[18] Cai Z W, Vasconcelos N. Cascade R-CNN: high quality object detection and instance segmentation[J]. IEEE Trans Pattern Anal Mach Intell, 2021, 43(5): 1483-1498. DOI:10.1109/TPAMI.2019.2956516.
[19] Fang Y X, Yang S S, Wang X G, et al. Instances as queries[C] //2021 IEEE/CVF International Conference on Computer Vision(ICCV). October 10-17, 2021, Montreal, QC, Canada. IEEE, 2021: 6890-6899. DOI:10.1109/ICCV48922.2021.00683.
[20] Wang X L, Zhang R F, KoNG T, et al. SOLOv2: Dynamic and fast instance segmentation[C]. ArXiv Comput Vis Pattern Recognit. 2020
[21] Du X Z, Lin T Y, Jin P C, et al. SpineNet: learning scale-permuted backbone for recognition and localization[C] //2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR). June 13-19, 2020. Seattle, WA, USA. IEEE, 2020: 11589-11598. DOI:10.1109/cvpr42600.2020.01161.
[22] Jocher G, Chaurasia A, StokEN A, et al. ultralytics/yolov5: v7.0-YOLOv5 SOTA realtime instance segmentation[J]. Zenodo, 2022.
[23] Wang C Y, Bochkovskiy A, Liao H 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(CVPR). June 17-24, 2023, Vancouver, BC, Canada. IEEE, 2023: 7464-7475. DOI:10.1109/CVPR52729.2023.00721.
[24] Vaswani A, Shazeer N, ParMAR N, et al. Attention is all you need[J]. Adv Neural Inf Process Syst, 2017, 5998-6008.
[25] Dai X Y, Chen Y P, Yang J W, et al. Dynamic DETR: end-to-end object detection with dynamic attention[C] //2021 IEEE/CVF International Conference on Computer Vision(ICCV). October 10-17, 2021, Montreal, QC, Canada. IEEE, 2021: 2968-2977. DOI:10.1109/ICCV48922.2021.00298.
[26] Han D C, Pan X R, Han Y Z, et al. FLatten transformer: vision transformer using focused linear attention[C] //2023 IEEE/CVF International Conference on Computer Vision(ICCV). October 1-6, 2023, Paris, France. IEEE, 2023: 5938-5948. DOI:10.1109/ICCV51070.2023.00548.
[27] Wang Y Q, Xu Z L, Wang X L, et al. End-to-end video instance segmentation with transformers[C] //2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR). June 20-25, 2021. Nashville, TN, USA. IEEE, 2021: 8737-8746. DOI:10.1109/cvpr46437.2021.00863.
[28] Rezatofighi H, Tsoi N, Gwak J, et al. Generalized intersection over union: a metric and a loss for bounding box regression[C] //2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR). June 15-20, 2019, Long Beach, CA, USA. IEEE, 2019: 658-666. DOI:10.1109/CVPR.2019.00075.
[29] Zheng Z H, Wang P, Liu W, et al. Distance-IoU loss: faster and better learning for bounding box regression[J]. Proc AAAI Conf Artif Intell, 2020, 34(7): 12993-13000. DOI:10.1609/aaai.v34i07.6999.
[30] Zheng Z H, Wang P, Ren D W, et al. Enhancing geometric factors in model learning and inference for object detection and instance segmentation[J]. IEEE Trans Cybern, 2022, 52(8): 8574-8586. DOI:10.1109/TCYB.2021.3095305.
[31] Zhang Y F, Ren W Q, Zhang Z, et al. Focal and efficient IOU loss for accurate bounding box regression[J]. Neurocomputing, 2022, 506: 146-157. DOI:10.1016/j.neucom.2022.07.042.
[32] Tong Z J, Chen Y H, Xu Z W, et al. Wise-IoU: bounding box regression loss with dynamic focusing mechanism[EB/OL]. 2023: arXiv: 2301.10051. https://doi.org/10.48550/arXiv.2301.10051.
[33] Wang C Y, Yeh I H, Liao H M. YOLOv9: learning what you want to learn using programmable gradient information[EB/OL]. 2024: ArXiv: 2402.13616. https://doi.org/10.48550/arXiv.2402.13616.
[34] Khanam R, Hussain M. YOLOv11: an overview of the key architectural enhancements[EB/OL]. 2024: arXiv: 2410.17725. https://doi.org/10.48550/arXiv.2410.17725. (