PMS-YOLO:一种用于车辆部件精确分割的多尺度特征融合注意力模型

PMS-YOLO:A multi-scale feature fusion attention model for precise segmentation of vehicle parts

摘要

为应对自动驾驶系统中车辆部件分割精度不足的挑战,提出了一种改进的多尺度特征融合注意力机制模型——PMS-YOLO。该模型通过替换原始主干网络为Ghost_HGNetV2,显著平衡了高效特征提取与计算效率,并利用深度金字塔结构和多阶段设计有效捕获多尺度上下文特征。在多尺度特征融合阶段,引入DySample模块进行细节对齐,减少特征上采样过程中信息丢失。同时,结合CPCA(Cross-patch contextual attention)模块提升高层特征处理能力,增强目标边界分割精度。在Carparts-seg数据集上的实验表明,PMS-YOLO相较基准模型YOLOv11n,mAP50_M提升了4.7%。并且在精度提升的同时,仍保持了优良的分割速度和小参数量。证明了PMS-YOLO的有效性。

Abstract

To address the challenge of insufficient segmentation accuracy of vehicle components in autonomous driving systems,an enhanced multi-scale feature fusion attention mechanism model named PMS-YOLO is proposed.The model significantly balances efficient feature extraction and computational efficiency by replacing the original backbone network with Ghost_HGNetV2,while effectively capturing multi-scale contextual features through a deep pyramid structure and multi-stage design.During the multi-scale feature fusion phase,a DySample module is introduced for detail alignment to mitigate information loss during feature upsampling.Concurrently,the integration of a CPCA (Cross-patch contextual attention) module enhances high-level feature processing capabilities and improves boundary segmentation accuracy.Experiments on the CarParts-seg dataset demonstrate that PMS-YOLO achieves a 4.7% improvement in mAP50_M, compared to the baseline model YOLOv11n.Notably,the model maintains excellent segmentation speed and compact parameter size while achieving higher accuracy,validating the effectiveness of PMS-YOLO.

关键词

车辆部件分割 / 多尺度特征融合 / YOLOv11 / 注意力机制 / 实例分割

Key words

vehicle component segmentation / multi-scale feature fusion / YOLOv11 / attention mechanism / instance segmentation

[1] Li Xinjin,Yang Yuanzhe,Yuan Yixiao,et al.Intelligent vehicle classification system based on deep learning and mul tisensor fusion[C]//Fifth International Conference on Computer Vision and Data Mining (ICCVDM 2024).B ellingham,WA:SPIE,2024:545-553.
[2] Pasupa K,Kittiworapanya P,Hongngern N,et al.Evaluation of deep learning algorithms for semantic segmentation of car parts[J].Complex and Intelligent Systems,2022,8(5):3613-3625.
[3] Ojha A,Sahu S P,Dewangan D K.Vehicle detection through instance segmentation using MaskR-CNN for intelligent vehicle system[C]//2021 5th International Conference on Intelligent Computing and Control Systems (ICICCS).Madurai,India:IEEE,2021:954-959.
[4] 韩烜宇,王安志,杨成帮,等.面向计算机视觉应用的扩散模型综述[J].贵州师范大学学报(自然科学版),2025,43(1):115-128.
[5] Pandharipande A,Cheng Chih-Hong,Dauwels J,et al.Sensing and machine learning for automotive perception:a review[J].IEEE Sensors Journal,2023,23(11):11097-11115.
[6] Fu Cebin,Tang Xiangyan,Yang Yue,et al.A survey of research progresses on instance segmentation based on deep learning[C]//International Conference on Big Data and Security.Singapore:Springer Nature Singapore,2023:138-151.
[7] 马俊燕,常亚楠.MFE-YOLOX:无人机航拍下密集小目标检测算法[J].重庆邮电大学学报(自然科学版),2024,36(1):128-135.
[8] Dubuisson M P,Jain A K.Contour extraction of moving objects in complex outdoor scenes[J].International Journal of Computer Vision,1995,14(1):83-105.
[9] Yang Ming,Fan Xiangyu.YOLOv8-Lite:a lightweight object detection model for real-time autonomous driving systems[J].IECE Transactions on Emerging Topics in Artificial Intelligence,2024,1(1):1-16.
[10] Gao Chengfei,Zhao Fengkui,Zhang Yong,et al.Research on multitask model of object detection and road segmentation in unstructured road scenes[J].Measurement Science and Technology,2024,35(6):065113.
[11] Zhao Xiaofeng,Zhang Wenwen,Zhang Hui,et al.ITD-YOLOv8:an infrared target detection model based on YOLOv8 for unmanned aerial vehicles[J].Drones,2024,8(4):161.
[12] Liu Wenze,Lu Hao,Fu Hongtao,et al.Learning to upsample by learning to sample[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision,Piscataway,NJ:IEEE,2023:6027-6037.
[13] Huang Hejun,Chen Zuguo,Zou Ying,et al.Channel prior convolutional attention for medical image segmentation[J].Computers in Biology and Medicine,2024,178:108784.
[14] Khanam R,Hussain M.YOLOv11:an overview of the key architectural enhancements[PP/OL].arXiv (2024-10-23)[2025-01-10].https://arxiv.org/abs/2410.17725.
[15] He Luhao,Zhou Yongzhang,Liu Lei,et al.Research and application of YOLOv11-Based object segmentation in intellig ent recognition at construction sites[J].Buildings,2024,14(12):3777.
[16] Jegham N,Koh C Y,Abdelatti M,et al.YOLO Evolution:a comprehensive benchmark and architectural review of YOLOv12,YOLO11,and their previous versions[PP/OL].arXiv (2024-11-01)[2025-01-15].https://arxiv.org/abs/2411.00201.
[17] Zhang Lei,Sun Zhipeng,Tao Hongjing,et al.Research on mine-personnel helmet detection based on multi-strategy-imp roved YOLOv11[J].Sensors,2024,25(1):170.
[18] Huang Zhanchao,Wang Jianlin,Fu Xuesong,et al.DC-SPP-YOLO:dense connection and spatial pyramid pooling based YOLO for object detection[J].Information Sciences,2020,522:241-258.
[19] Lu Yong,Sun Minghao.Lightweight multidimensional feature enhancement algorithm LPS-YOLO for UAV remote sensing target detection[J].Scientific Reports,2025,15(1):1340.
[20] Wang Zhou,Su Yuting,Kang Feng,et al.PC-YOLOv11s:a lightweight and effective feature extraction method for small target image detection[J].Sensors,2025,25(2):348.
[21] Sapkota R,Ahmed D,Karkee M.Comparing YOLOv8 and Mask R-CNN for instance segmentation in complex or chard environments[J].Artificial Intelligence in Agriculture,2024,13:84-99.
[22] Zhai Haozhou,Du Jinwei,Ai Yuhui,et al.Edge deployment of deep networks for visual object detection:a review[J].IEEE Sensors Journal,2024,25(11):18662-18683.
[23] 王磊,张斌,吴奇鸿.RCSA-YOLO:改进YOLOv8的SAR舰船实例分割[J].计算机工程与应用,2024,60(18):103-113.
[24] 赵南南,高翡晨.基于改进YOLOv8的交通场景实例分割算法[J].计算机工程,2025,51(1):198-207.
[25] 谭旭,赵骥.改进YOLOv8的汽车表面伤损实例分割模型[J].计算机工程与应用,2024,60(14):197-208.
[26] Hidayatullah P,Syakrani N,Sholahuddin M R,et al.YOLOv8 to YOLO11:a comprehensive architecture in-depth comparative review[PP/OL].arXiv (2025-01-22)[2025-01-10].https://arxiv.org/abs/2501.13400.
[27] Chen Hao,Sun Kunyang,Tian Zhi,et al.Blendmask:top-down meets bottom-up for instance segmentation[C]//Proc eedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,IEEE,2020:8573-8581.
[28] Hasan S M K,Linte C A.U-NetPlus:a modified encoder-decoder U-Net architecture for semantic and in stance segmentation of surgical instruments from laparoscopic images[C]//2019 41st Annual Internatio nal Conference of the IEEE Engineering in Medicine and Biology Society (EMBC),Berlin,Germany:IEEE,2019:7205-7211.
[29] Yan Zengqiang,Yang Xin,Cheng Kwang-Ting.Enabling a single deep learning model for accurate gland instance segmentation:a shape-aware adversarial learning framework[J].IEEE Transactions on Medical Imaging,2020,3(6):2176-2189.
[30] Alif M A R.YOLOv11 for vehicle detection:advancements,performance,and applications in intelligent transportation systems[PP/OL].arXiv preprint arXiv(2024-10-28)[2025-03-11].https://arxiv.org/abs/2410.22898.
[31] Lin Zhenyu,Yun Bensheng,Zheng Yanan.LD-YOLO:a lightweight dynamic forest fire and smoke detection model with dysample and spatial context awareness module[J].Forests,2024,15(9):1630.
[32] Russo G.Car-seg Dataset[EB/OL].(2023-11-01)[2024-01-24].https://universe.roboflow.com/gianmarco-russo-vtqxr/car-seg-un1pm.
[33] Factorypackage.factory_package dataset[EB/OL].(2024-01-01)[2024-01-24].https://universe.roboflow.com/factorypackage/factory_package.
[34] Car.Car parts detection Dataset[EB/OL].(2024-06-01)[2025-04-22].https://universe/roboflow.com/car-2kcsx/car-parts-detection-mjthe.

基金资助

国家自然科学基金项目(62161019)