河北水利电力学院学报

河北水利电力学院学报 ›› 2025, Vol. 35 ›› Issue (2): 35-41.DOI: 10.16046/j.cnki.issn2096-5680.2025.02.007

• 人工智能与机器人 • 上一篇    下一篇

便携收费终端条件下改进型车牌识别算法研究

吉庆昌1,2, 刘强3, 栗梦媛1, 孙东来1,2, 王继超1   

  1. 1.河北水利电力学院电气自动化系,河北省沧州市黄河西路49号 061001;
    2.河北省高校交通基础设施数智化应用技术研发中心,河北省沧州市黄河西路49号 061001;
    3.河北省公路事业发展中心,河北省石家庄市长安区裕华东路509号 050032
  • 收稿日期:2025-02-13 修回日期:2025-03-18 出版日期:2025-06-30 发布日期:2025-07-04
  • 通讯作者: 刘 强(1982-),男,天津人,正高级工程师,主要研究领域为高速公路智能化应用。E-mail:45936960@qq.com
  • 作者简介:吉庆昌(1982-),男,河北献县人,硕士,主要研究领域为图像处理、故障诊断。E-mail:jiqingchang@hbwe.edu.cn
  • 基金资助:
    河北省高层次人才资助项目(A202001041)

Research on Improved License Plate Recognition Algorithm under the Condition of Portable Toll Collection Terminal

JI Qingchang1,2, LIU Qiang3, LI Mengyuan1, SUN Donglai1,2, WANG Jichao1   

  1. 1. Hebei University of Water Resources and Electric Engineering,061001,Cangzhou,Hebei,China;
    2. Hebei Higher Institute of Transportation Infrastructure Research and Development Center for Digital and Intelligent Technology Application,061001,Cangzhou,Hebei,China;
    3. Hebei Highway Development Center, 050011, Shijiazhuang, Hebei, China
  • Received:2025-02-13 Revised:2025-03-18 Online:2025-06-30 Published:2025-07-04

PDF

摘要： 为了解决高速收费车道设备发生故障或车流量较大时出现的车辆拥堵问题,文中提出了一种基于YOLOv5模型的改进型端到端的轻量化车牌识别算法。该算法首先将轻量级网络FasterNet作为YOLOv5主干模块,有效降低了模型计算复杂度;其次融入ECA-Net注意力机制,使得模型能够更快地定位至候选区域;使用EIoU作为损失函数,有利于小目标识别。实验结果表明改进后的车牌识别算法相对于传统基于YOLOv5模型的车牌识别算法,mAP值提高了0.8%,FLOPs降低了18.9%,更适合部署在计算能力相对较低的便携收费终端。将开发的改进型车牌识别算法部署在便携收费终端条件下进行测试,结果验证了算法的有效性。

关键词: 便携收费终端, 车牌识别, YOLOv5, 模型轻量化, 注意力机制, 损失函数

Abstract: An improved end-to-end lightweight license plate recognition algorithm based on YOLOv5 was proposed to address vehicle congestion caused by equipment failures or heavy traffic at highway toll stations. The algorithm was specifically designed for deployment on portable toll collection terminals with limited hardware resources to enable mobile manual charging/card issuing operations. Three key modifications were implemented in the YOLOv5 architecture. Firstly, the backbone network was replaced with FasterNet to effectively reduce computational complexity. Secondly, the ECA-Net attention mechanism was incorporated to enhance candidate region localization efficiency. Finally, the EIoU loss function was utilized to improve small target recognition accuracy. Experimental results demonstrated that the modified algorithm achieved 0.8% higher mAP and 18.9% lower FLOPs compared to the original YOLOv5-based license plate recognition model. Comprehensive field tests were conducted under actual portable toll collection terminal deployment conditions. The effectiveness of the algorithm was verified through practical application scenarios.

Key words: portable toll collection terminal, license plate recognition, YOLOv5, model lightweighting, attention mechanism, loss function

中图分类号: