1.China Mobile Communications Group Zhejiang Co. ,Ltd. ,Hangzhou 310016,Zhejiang,China
2.School of Electronic and Electrical Engineering,Shanghai University of Engineering Science,Shanghai 201620,China
3.Engineering Research Center of Intelligent Technology for Agriculture,Ministry of Education,Huazhong Agricultural University,Wuhan 430070,Hubei,China
4.Key Laboratory of Smart Farming Technology for Agricultural Animals,Ministry of Agriculture and Rural Affairs,College of Informatics,Huazhong Agricultural University,Wuhan 430070,Hubei,China
Show less
文章历史+
Received
Published
2025-05-09
2026-02-24
Issue Date
2026-07-23
PDF (1842K)
摘要
目标检测是计算机视觉的基石,在医疗成像、工业缺陷检测、自动驾驶和农业监测等领域有着广泛的应用。YOLO(You Only Look Once)作为单阶段目标检测的开创性算法,以其卓越的实时性和高效性成为计算机视觉领域的核心技术。从2016年YOLOv1的首次亮相到2025年YOLOv12的发布,YOLO通过持续优化网络架构、改进损失函数设计和创新数据增强策略,显著提升了检测性能和应用适配性。本文系统回顾了YOLO长达十年的发展历程,分析了其在不同领域的应用,并探讨了其快速发展背后的驱动力,包括深度学习技术的突破、开源社区的广泛协作以及对实时性和边缘计算的迫切需求。与双阶段算法(如Faster R-CNN)和基于Transformer的方法(如DETR)相比,YOLO在速度和工程部署上展现出显著优势,但在小目标检测、密集场景处理和复杂背景下的鲁棒性方面仍面临挑战。未来YOLO或可通过整合轻量级Transformer模块、优化多任务学习框架以及引入生成式AI增强数据多样性,有望进一步提升精度和场景适应能力。
Abstract
Target detection is the cornerstone of computer vision, with a wide range of applications such as medical imaging, industrial defect detection, autonomous driving and agricultural monitoring, etc. YOLO (You Only Look Once), a pioneering algorithm for single-stage target detection, has become a core technology in the field of computer vision due to its excellent real-time performance and high efficiency. From the debut of YOLOv1 in 2016 to the release of YOLOv12 in 2025, YOLO has significantly improved its detection performance and application suitability through continuous optimization of the network architecture, improvement of the loss function design and innovative data enhancement strategies. In this paper, we systematically review YOLO’s decade-long development history, analyze its applications in different domains, and discuss the driving forces behind its rapid development, including breakthroughs in deep learning technology, extensive collaboration in the open source community, and the pressing need for real-time and edge computing. Compared with two-stage algorithms (e.g., Faster R-CNN) and Transformer-based methods (e.g., DETR), YOLO demonstrates significant advantages in terms of speed and engineering deployment, but its limitation is that it still faces challenges in terms of small-target detection, dense-scene processing, and robustness in complex contexts. In the future, YOLO may be able to enhance data diversity by integrating a lightweight Transformer module, optimizing the multi-task learning framework, and introducing generative AI, which is expected to further improve accuracy and scene adaptation.
目标检测作为计算机视觉的核心技术,旨在从图像或视频中识别并定位物体,广泛应用于医疗影像分析、工业缺陷检测、自动驾驶和农业监测等领域。传统目标检测方法依赖手工设计的特征提取,受限于表达能力[1],难以应对复杂场景。2012年,AlexNet引入卷积神经网络(CNN)[2],凭借强大的特征学习能力开启了深度学习时代,显著提升了目标检测的性能。此后,目标检测算法逐渐分为双阶段和单阶段两大流派[3]。双阶段方法(如R-CNN系列)通过候选区域生成和精细回归实现高精度,但计算复杂度高,难以满足实时需求。相比之下,单阶段方法以速度见长,尤其以YOLO(You Only Look Once)为代表,通过一次前向传播同时预测边界框和类别概率,树立了实时检测的里程碑。
YOLOv1(You Only Look Once)由Redmon等于2016年提出[4],受到GoogleNet[11]和Network in Network的启发,使用1×1的卷积层减少特征图数量并保持较低参数,将输入图像划分为S×S个网格,每个网格产生B个候选框和C个类别概率,最后输出一个多维数组包含上述信息。在VOC 2007数据集[12]上实现了63.4%的mAP(平均精度)。
在数据处理方面,YOLOv4引入了许多特殊的训练技巧,提出了BoF策略,其中比较出名的便是参考CutMix方法的一种新的数据增强技术[5]:Mosaic。它通过将四张不同的图片拼接在一起,形成一张新的训练图像。由于拼接后的图像使模型在训练时看到了更多样化的背景和目标,这种方式不仅增强了模型的鲁棒性,还在一定程度上缓解了数据集不足的问题。YOLOv4还采用了CIoU损失函数(Complete Intersection over Union Loss),考虑了边界框的中心点距离、面积比和长宽比等因素,从而更全面地评估预测框和真实框之间的差异,提高了边界框回归的精度。
KITTI数据集:KITTI是一个专注于自动驾驶的视觉数据集,由卡尔斯鲁厄理工学院(KIT)和丰田技术研究院(Toyota Technological Institute at Chicago,TTI-C)合作创建。它包含多种任务的标注,如目标检测、立体视觉、光流等。该数据集主要关注自动驾驶场景中的物体,如车辆、行人、骑行者等,注释信息包括2D边界框和3D边界框。上述数据集为YOLO模型的训练和评估提供了数据支持,其多样性直接影响评价指标。以下将介绍YOLO常用的评估指标及其计算方法。
LÜL, CHENGH, ZHUH T, et al. Progress of research and application of object detection based on deep learning[J]. Electronics & Packaging, 2022, 22(1): 72-80. DOI: 10.16257/j.cnki.1681-1070.2022.0114 (Ch ).
[5]
REDMONJ, DIVVALAS, GIRSHICKR, et al. You only look once: Unified, real-time object detection[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). New York: IEEE Press, 2016: 779-788. DOI: 10.1109/CVPR.2016.91 .
MAOS H, WANGW D. Research review of YOLO series object detection algorithms based on deep learning[J]. Journal of Yan’an University (Natural Science Edition), 2024, 43(2): 88-95. DOI: 10.13876/J.cnki.ydnse.230044 (Ch ).
MIZ, LIANZ. Review of YOLO methods for universal object detection[J]. Computer Engineering and Applications, 2024, 60(21): 38-54. DOI: 10.3778/j.issn.1002-8331.2404-0130 (Ch ).
[10]
WANGL L, LIJ J, KANGF. Crack location and degree detection method based on YOLOX model[J]. Applied Sciences, 2022, 12(24): 12572. DOI: 10.3390/app122412572 .
[11]
WANGC, YEHI, LIAOH M. You only learn one representation: Unified network for multiple tasks[J]. Journal of Information Science and Engineering, 2023, 39(3): 691-709. DOI: 10.6688/JISE.202305_39(3).0015 .
[12]
SIMONYANK, ZISSERMANA. Very deep convolutional networks for large-scale image recognition[C]//International Conference on Learning Representations. Sam Diego: Computational and Biological Learning Society, 2015.DOI:10.48550/arXiv.1409.1556 .
[13]
WANGC Y, BOCHKOVSKIYA, LIAOH M. Scaled-YOLOv4: Scaling cross stage partial network[C]//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). New York: IEEE Press, 2021: 13024-13033. DOI: 10.1109/CVPR46437.2021.01283 .
[14]
SZEGEDYC, LIUW, JIAY Q, et al. Going deeper with convolutions[C]//2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). New York: IEEE Press, 2015: 1-9. DOI: 10.1109/CVPR.2015.7298594 .
[15]
EVERINGHAMM, VAN GOOLL, WILLIAMSC K I, et al. The pascal visual object classes (VOC) challenge[J]. International Journal of Computer Vision, 2010, 88(2): 303-338. DOI: 10.1007/s11263-009-0275-4 .
[16]
YUJ H, JIANGY N, WANGZ Y, et al. UnitBox: An advanced object detection network[C]//Proceedings of the 24th ACM International Conference on Multimedia. New York: ACM, 2016: 516-520. DOI: 10.1145/2964284.2967274 .
[17]
REDMONJ, FARHADIA. YOLO9000: Better, faster, stronger[C]//2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). New York: IEEE Press, 2017: 6517-6525. DOI: 10.1109/CVPR.2017.690 .
[18]
IOFFES, SZEGEDYC. Batch normalization: Accelerating deep network training by reducing internal covariate shift[C]//Proceedings of the 32nd International Conference on International Conference on Machine Learning. Lille:PMLR,2015: 448-456. DOI: 10.5555/3045118.3045167 .
[19]
LINT Y, MAIREM, BELONGIES, et al. Microsoft COCO: Common objects in context[C]//Computer Vision — ECCV 2014. Cham: Springer, 2014: 740-755. DOI: 10.1007/978-3-319-10602-1_48 .
[20]
LIZ, YANGL, ZHOUF.FSSD: Feature fusion single shot multibox detector[DB/OL].[2025-01-03].
HEK, ZHANGX, RENS, et al.Spatial pyramid pooling in deep convolutional networks for visual recognition[J].IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015, 37(9): 1904-1916.DOI:10.1109/TPAMI.2015.2389824 .
[23]
MIKOLOVT, KOMBRINKS, BURGETL, et al. Extensions of recurrent neural network language model[C]//2011 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). New York: IEEE Press, 2011: 5528-5531. DOI: 10.1109/ICASSP.2011.5947611 .
[24]
BOCHKOVSKIYA, WANGC Y, LIAOH M. YOLOv4: Optimal speed and accuracy of object detection[EB/OL]. 2020: arXiv: 2004.10934. DOI: 10.48550/arXiv.2004.10934 .
[25]
LIUS, QIL, QINH F, et al. Path aggregation network for instance segmentation[C]//2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2018: 8759-8768. DOI: 10.1109/CVPR.2018.00913 .
[26]
WANGC Y, LIAOH Y M, WUY H, et al. CSPNet: A new backbone that can enhance learning capability of CNN[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). New York: IEEE Press, 2020: 1571-1580. DOI: 10.1109/CVPRW50498.2020.00203 .
[27]
YUNS, HAND, CHUNS, et al. CutMix: Regularization strategy to train strong classifiers with localizable features[C]//2019 IEEE/CVF International Conference on Computer Vision (ICCV). New York: IEEE Press, 2019: 6022-6031. DOI: 10.1109/ICCV.2019.00612 .
[28]
LIC Y, LIL L, JIANGH L, et al. YOLOv6: A single-stage object detection framework for industrial applications[EB/OL]. 2022: arXiv: 2209.02976.
[29]
WANGC Y, BOCHKOVSKIYA, LIAOH 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). New York: IEEE Press, 2023: 7464-7475. DOI: 10.1109/CVPR52729.2023.00721 .
[30]
SOHANM, SAIRAM T, RAMI REDDYC V. A review on YOLOv8 and its advancements[C]//Data Intelligence and Cognitive Informatics. Singapore: Springer, 2024: 529-545. DOI: 10.1007/978-981-99-7962-2_39 .
[31]
WANGC Y, YEHI H, MARK LIAOH Y. YOLOv9: Learning what you want to learn using programmable gradient information[C]//Computer Vision — ECCV 2024. Cham: Springer, 2025: 1-21. DOI: 10.1007/978-3-031-72751-1_1 .
[32]
WANGA, CHENH, LIUL, et al. Yolov10: Real-time end-to-end object detection[J]. Advances in Neural Information Processing Systems, 2024, 37:107984-108011.DOI:10.48550/arXiv.2405.14458 .
[33]
KHANAMR, HUSSAINM. YOLOv11: An overview of the key architectural enhancements[EB/OL]. 2024: arXiv: 2410.17725. DOI: 10.3390/solar5010006 .