1 Yanzhao Electric Power Laboratory, North China Electric Power University, Baoding 071003, China
2 Department of Automation, North China Electric Power University, Baoding 071003, China
3 Baoding Key Laboratory of Intelligent Robot Perception and Control in Electric Power System, Baoding 071003, China
4 Electric Power College, South China University of Technology, Guangzhou 510000, China
5 Engineering Training and Innovation and Entrepreneurship Education Center, North China Electric Power University, Baoding 071003, China
Show less
文章历史+
Received
Published
Revised
2025-08-28
2026-08-08
2026-01-09
Online Date
2026-09-03
PDF (11785K)
摘要
为了有效解决车辆损伤中因类间特征相似和类内多样性所导致的漏检和误检问题,提出了一种融合显性细节特征推理的车辆损伤检测方法。首先,将车辆损伤的显性细节边缘特征引入WIoU(wise intersection over union)框架,形成基于车辆损伤熵的WIoU(vehicle damage WIoU, V-WIoU)。V-WIoU通过学习不同损伤的显性边缘曲率,计算曲率分布熵以衡量掩码边界复杂度,并将其融合到WIoU的离群度项,从而实现对损伤类型学习难度差异的有效区分,增强了模型的特征提取能力。其次,设计了细节特征捕捉模块,用于高效提取车辆显性细节纹理与空间显性特征信息;在此基础上进一步提出了细节特征逻辑校正模块,并融合进车辆损伤检测模型中。该模块通过引入逻辑校正规则,对显性特征分布进行推理性约束,有效缓解了损伤检测中的误检问题。实验在自建的10类车辆损伤数据集上进行,结果表明所提出模型在总体检测与分割性能上均优于其他先进实例分割模型。与基线模型相比,本文方法在目标框检测准确率和分割准确率上分别提升了6.8%和7.6%。消融实验进一步验证了各模块对检测性能的显著贡献。综合来看,本文模型通过V-WIoU、细节特征捕捉模块与细节特征逻辑校正模块的协同作用,有效缓解了特征提取不足、类间相似与类内多样性带来的检测难题,显著提升了车辆损伤检测的精度与鲁棒性。
Abstract
To effectively address the issues of missed and false detections caused by inter-class feature similarity and intra-class diversity in vehicle damage detection, a vehicle damage detection method incorporating explicit detail feature reasoning was proposed. Firstly, the explicit detail edge features of vehicle damage are introduced into WIoU (wise intersection over union) framework, forming V-WIoU (vehicle damage WIoU) based on vehicle damage entropy. V-WIoU was designed to learn explicit edge curvatures of different damage types. Subsequently, curvature distribution entropy was calculated to quantify the complexity of mask boundaries. This entropy metric was integrated into the outlier term of WIoU, which enables effective differentiation of learning difficulties among various damage categories and enhances the feature extraction capability of the model. Secondly, a detail feature capture module was developed to efficiently extract explicit detail textures and spatial explicit feature information of vehicles. On this basis, a detail feature logical correction module was further proposed and embedded into the vehicle damage detection model. This module introduces logical correction rules to impose reasoning constraints on the distribution of explicit features, thereby effectively alleviating false detection issues in damage detection tasks. Experiments are conducted on a self-constructed dataset containing 10 types of vehicle damages, and the results demonstrate that the proposed model outperforms other advanced instance segmentation models in terms of overall detection and segmentation performance. Compared with the baseline model, the proposed method increases the accuracy of object bounding box detection and mask segmentation by 6.8% and 7.6%, respectively. Ablation experiments further verify the significant contributions of each proposed module to the improvement of detection performance. In conclusion, through the synergistic effects of V-WIoU, the detail feature capture module, and the detail feature logical correction module, the proposed model effectively mitigates the detection challenges caused by insufficient feature extraction, inter-class similarity, and intra-class diversity, thus significantly enhancing the accuracy and robustness of vehicle damage detection.
车辆损伤检测任务的应用包括损伤分类[5-6]和损伤检测。损伤检测需要在损伤分类的基础上对损伤进行精确定位,这对检测模型提出了更高的要求。Peng等[7]提出了一种多视角汽车损伤检测模型(multi-view car damage detection model,MVA-CDD),通过特征分割、融合和对齐模块实现全局与局部特征的有效整合,显著提升了检测精度。Setyawan等[8]系统地比较了单阶段和双阶段检测模型在车辆损伤检测场景中的性能,验证了YOLO(you only look once)轻量化模型的应用潜力。彭锦波[9]从多视角深度学习的角度去解析并检测车辆损伤,一定程度上解决了传统检测技术中视角限制的问题。然而,上述研究主要集中于有限数量的车辆部件和特定类型的损伤。这些工作对通用深度学习模型的依赖,往往导致其难以有效提取细微的损伤特征,进而造成在实际车辆损伤检测场景中漏检和误检问题频发。鉴于此,提升模型对车辆损伤的特征提取能力和对损伤细节的挖掘能力仍然是该领域一个至关重要的研究目标。
LiFeng, DaiZhong, HuiDapeng, et al. Implementation and effect of highway construction vehicle detection using the improved YOLOv5 method[J]. Science Technology and Engineering, 2025, 25(27): 11842-11850.
ZhaiYongjie, ZhuHaoshuo, LiJiawei, et al. VDD-YOLO: evalua-tion method for vehicle deformation damage[J]. Science Technology and Engineering, 2025, 25(29): 12560-12569.
[5]
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.
[6]
XieDongsheng. Research on intelligent vehicle loss assessment algorithm based on deep learning[D]. Tianjin: Tianjin University, 2019.
[7]
SinghR, AyyarMP, PavanT V S, et al. Automating car insurance claims using deep learning techniques[C]// IEEE Fifth Internatio-nal Conference on Multimedia Big Data (BigMM). Singapore: IEEE, 2019: 199-207.
[8]
TianX, HanH. Deep convolutional neural networks withtransfer learning for automobile damage image classification[J]. Journal of Database Management, 2022, 33(3): 1-16.
[9]
DongZ, WuY, PeiM, et al. Vehicle type classification using a semisupervised convolutional neural network[J]. IEEE Transactions on Intelligent Transportation Systems, 2015, 16(4): 2247-2256.
[10]
PengJ, DongS, YuanH, et al. Car damage detection based on multi-view fusion and alignment: dataset and method[J]. IEEE Transactions on Intelligent Transportation Systems, 2025, 26(4): 4717-4730.
[11]
SetyawanH A, BustamamA, BuyungR A. Analysis performance of one-Stage and two stage object detection method for car damage detection[J]. International Journal of Advanced Computer Science & Applications, 2024, 15(7): 3-6.
[12]
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.
[13]
PengJinbo. Research on vehicle damage detection based on multi-view deep learning[D]. Guangzhou: South China University of Technology, 2023.
[14]
HeK, GkioxariG, DollárP, et al. Mask R-CNN[C]// Proceedings of the IEEE International Conference on Computer Vision. Venice: IEEE, 2017: 2961-2969.
[15]
BolyaD, ZhouC, XiaoF, et al. Yolact: real-time instance segmentation[C]// Proceedings of the IEEE/CVF International Conference on Computer Vision. Seoul: IEEE, 2019: 9157-9166.
[16]
WangC Y, BochkovskiyA, LiaoH Y M. YOLOv7: trainable bag-of-freebies sets new state-of-the-art for real-time object detectors[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Vancouver: IEEE, 2023: 7464-7475.
[17]
CaoL, ZhengX, FangL. The semantic segmentation of standing tree images based on the YOLO V7 deep learning algorithm[J]. Electronics, 2023, 12(4): 929.
[18]
WangC Y, YehI H, LiaoH Y M. YOLOv9: learning what you want to learn using programmable gradient information[J]. arXiv PrePrint, arXiv: 2402.13616, 2024.
SunHong, SunJinwang, LeiJingfa, et al. Initial estimation of digital image correlation deformation based on improved SIFT feature matching[J]. Journal of Ordnance Equipment Engineering, 2025, 46(5): 256-265, 312.
LiuYarong, ZhiZhengxin, XieXiaolan. Rolling bearing fault feature extraction method based on improved VMD and entropy value feature fusion[J]. Science Technology and Engineering, 2025, 25(28): 12013-12022.
[25]
ChaiT, XiaoZ, ShenX, et al. Transdeep: transformer-integrated deeplabv3+ for image semantic segmentation[J]. IEEE Access, 2025, 13: 6277-6291.
[26]
AshishV. Attentionis all you need[J]. Advances in Neural Information Processing Systems, 2017, 30: 6000-6010.
[27]
RezatofighiH, TsoiN, GwakJ Y, et al. Generalized intersection over union: a metric and a loss for bounding box regression[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Long Beach: IEEE, 2019: 658-666.
[28]
ZhengZ, WangP, LiuW, et al. Distance-IoU loss: faster and better learning for bounding box regression[C]// Proceedings of the AAAI Conference on Artificial Intelligence. New York: AAAI, 2020: 12993-13000.
[29]
ZhengZ, WangP, RenD, et al. Enhancing geometric factors in model learning and inference for object detection and instance segmentation[J]. IEEE Transactions on Cybernetics, 2021, 52(8): 8574-8586.
[30]
ZhangY F, RenW, ZhangZ, et al. Focal and efficient IOU loss for accurate bounding box regression[J]. Neurocomputing, 2022, 506: 146-157.
[31]
DuS, ZhangB, ZhangP. Scale-sensitive IOU loss: an mproved regression loss function in remote sensing object detection[J]. IEEE Access, 2021, 9: 141258-141272.
[32]
TongZ, ChenY, XuZ, et al. Wise-IoU: bounding box regression loss with dynamic focusing mechanism[J]. arxiv Preprint arxiv: 2301.10051, 2023.
[33]
HeK, ZhangX, RenS, et al. Deep residual learning for image recognition[C]// Proceedings of the IEEE Conference on Compu-ter Vision and Pattern Recognition. Las Vegas: IEEE, 2016: 770-778.
[34]
TanM, PangR, LeQ V. Efficientdet: scalable and efficient object detection[C]// Proceedings of the IEEE/CVF Conference onComputer Vision and Pattern Recognition. Online: IEEE, 2020: 10781-10790.
[35]
CaiZ, VasconcelosN. Cascade R-CNN: high quality object detection and instance segmentation[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019, 43(5): 1483-1498.
[36]
HuangZ, HuangL, GongY, et al. Mask scoring R-CNN[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Long Beach: IEEE, 2019: 6409-6418.
[37]
WangX, ZhangR, KongT, et al. Solov2: dynamic and fastinstance segmentation[J]. Advances in Neural Information Processing Systems, 2020, 33: 17721-17732.
[38]
FangY, YangS, WangX, et al. Instances as queries[C]// Proceedings of the IEEE/CVF International Conference on Computer Vision. Montreal: IEEE, 2021: 6910-6919.
[39]
ChenK, PangJ, WangJ, et al. Hybrid task cascade for instance segmentation[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Long Beach: IEEE, 2019: 4974-4983.
[40]
KirillovA, WuY, HeK, et al. Pointrend: Image segmentation as rendering[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Online: IEEE, 2020: 9799-9808.