Navigation College, Dalian Martime University, Dalian 116026, China
Show less
文章历史+
Received
Published
Revised
2025-09-08
2026-08-18
2026-01-12
Online Date
2026-09-04
PDF (12201K)
摘要
危险货物集装箱标志的精准识别对于航运安全监管、人员防护及物流智能化具有重要意义。为提高标志识别精度,基于改进的YOLOv11(you only look once version 11)算法,提出一种双阶段端到端的图文融合识别框架。首先,在YOLOv11主干网络基础上引入CBAM(convolutional block attention module)注意力机制,采用基于颜色空间的Mosaic-Focus数据增强策略,建立Char-Focal-EIoU(efficient intersection over union)损失函数以提升小目标检测与字符定位精度,并在后处理阶段运用DT-NMS(dual-threshold non-maximum suppression)策略增强低置信度字符目标的召回率。其次,根据改进YOLOv11算法的目标检测结果,借助Paddle OCR光学字符识别模块,提升图文识别系统的字符结构化解析能力及对危险货物标志中UN编号等关键字符的识别精度。最终实验结果表明,基于该算法构建的危险货物标识识别系统在多个关键性能指标上均优于主流YOLO系列模型。改进算法显著增强了系统在复杂场景下的视觉感知能力,使标志检测精度、字符结构解析效果及系统响应效率得到全面提升,从而验证了双阶段端到端图文融合框架的先进性与实用性。
Abstract
Accurate recognition of dangerous goods container signs is essential for maritime safety supervision, personnel protection, and logistics intelligence. Based on an enhanced YOLOv11(you only look once version 11) model, a two-stage end-to-end image-text fusion framework was created to increase recognition performance in complicated situations. In the initial stage, CBAM(convolutional block attention module), a color-space-based Mosaic-Focus method, and the Char-Focal-EIoU regression loss were combined into the backbone, while DT-NMS (dual-threshold non-maximum suppression) was employed in post-processing to boost small-target localization and improve the recall of low-confidence character instances. Paddle OCR was used in the second step to carry out structured character parsing, which made it possible to recognize important identifiers like UN numbers with greater accuracy. Experimental findings indicate that the proposed framework exceeds standard YOLO models in detection accuracy, character identification, and system responsiveness,demonstrate enhanced visual perception in challenging circumstances and confirm the better performance and usefulness of the suggested two-stage image-text fusion technique for identifying dangerous goods container signs.
近年来,目标检测与文字识别技术发展迅速,深度学习方法凭借其提取图像特征快速、检测精度高、泛化能力强等优点,已经在图像检测领域广泛使用[2]。单阶段检测算法的代表YOLO(you only look once)系列凭借其强大的特征提取能力和一步推理的时间优势展现出卓越的性能[3-7]。袁小平等[8]采用YOLOv4算法将卷积模块与自注意力机制融合以提高复杂环境下的目标识别精度。刘桐等[9]将YOLOv5+ECA(efficient channel attention)注意力机制模块引入交通标志检测中,将准确率和平均精度分别提高2.4%及0.4%。郭宇翔[10]提出了YOLOv8改进算法SK-SG YOLOv8(selective kernel and SPD-Conv with GhostNet based YOLOv8),通过引入SPD-Conv(space-to-depth convolution)模块与Neck网络中的SK选择性核注意力机制,实现检测速度的优化。宋晓茹等[11]提出改进的YOLOv11算法,通过引入深度可分离卷积DWConv(depthwise convolution),完成复杂交通场景下车辆检测,将检测精度提升。
选择将CBAM模块添加到YOLOv11颈部网络的特征融合层之后,能够在不同尺度的特征图上同时应用通道注意力和空间注意力[20],进一步提升特征图的表达能力。实验表明,CBAM模块能有效提升模型对小目标、类间相似目标的识别能力,平均检测精度(mean average precision,mAP)获得明显提升。如图4所示为CBAM模块在YOLOv11网络结构中的位置图。
MoleroG D, SantarremigiaF E, Aragonés-BeltránP, et al. Total safety by design: increased safety and operability of supply chain of inland terminals for containers with dangerous goods[J]. Safety Science, 2017, 100: 168-182.
KuangXianyan, WangXingxing, WangLongfeng, et al. Multi scale aggregated remote sensing image small object detection algorithm based on improved YOLO[J]. Science Technology and Engineering, 2025, 25 (20): 8560-8570.
[4]
RedmonJ, DivvalaS, GirshickR, et al. You only look once: unified, real-time object detection[C]// IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas: IEEE, 2016: 779-788.
[5]
MujahidA, AwanM J, YasinA, et al. Real-time hand gesture recognition based on deep learning YOLOv3 model[J]. Applied Sciences, 2021, 11(9): 4164.
[6]
QiuZ, ZhuX, LiaoC, et al. Detection of transmission line insulator defects based on an improved lightweight YOLOv4 model[J]. Applied Sciences, 2022, 12(3): 1207.
WangKan, WangHao, LiuXiaolei, et al. Research on safety management of dangerous goods container operations in smart port shipping[J]. Safety, 2024, 45(9): 77-85.
[21]
KhanamR, HussainM. YOLOv11: an overview of the key architectural enhancements[J]. arXiv Preprint, arXiv: 2410. 17725, 2024.
[22]
HsiehC C, HsuC H, HuangW H. A two-stage road sign detection and text recognition system based on YOLOv7[J]. Internet of Things, 2024, 27: 101330.
[23]
AlifM A R. YOLOv11 for vehicle detection: advancements, performance, and applications in intelligent transportation systems[J]. arXiv Preprint, arXiv: 2410. 22898, 2024.
AnMengjun, WuChaoming, DengChengzhi. MEC-YOLOv11n: detection algorithm for small floating objects on water surface[J]. Electronic Measurement Technology, 2025, 48(14): 140-146.
SongJianhui, XiaBin, ZhaoYawei, et al. License plate recognition algorithm based on improved YOLOv8 and LPRNet[J]. Journal of Shenyang University of Technology, 2025, 44(3): 39-46.
Changsha University of Science and Technology. Resolution feature fusion extraction method and recognition method for multi pulse radar signals: 202110028601. 3[P]. 2021-03-26.
FanNan, XiaoShibin, WangHuanpeng, et al. Patent reference mark detection based on improved EAST algorithm[J]. Journal of Beijing University of Information Science and Technology (Natural Science Edition), 2021, 36(4): 95-100.
YangJingkai, LiuWeiyu, WangJingyuan. Research on deep learning based classification of toll vehicles on highways[J]. Internet of Things Technology, 2024, 14(2): 123-126.
LiXuewei, MaoXiaokun, SunBin, et al. Small target detection algorithm for UAV aerial photography based on YOLOv11n[J/OL]. (2025-02-10)[2026-01-12]. https://link.cnki.net/urlid/42.1266.TN.20250210.1015.002.