一种改进多尺度融合的电动汽车充电口识别方法  

An improved multi-scale fusion identification method for electric vehicle charging ports

作　　者：赵晓东[1] 刘瑞庆 王向[1] 温士涛 ZHAO Xiaodong;LIU Ruiqing;WANG Xiang;WEN Shitao(School of Information Science and Engineering,Hebei University of Science and Technology,Shijiazhuang 050018,China;93507 Unit,People’s Liberation Army of China,Shijiazhuang 050200,China)

机构地区：[1]河北科技大学信息科学与工程学院,石家庄050018 [2]中国人民解放军93507部队,石家庄050200

出　　处：《重庆理工大学学报（自然科学）》2024年第7期118-126,共9页Journal of Chongqing University of Technology：Natural Science

基　　金：河北省高等学校科学技术重点研究项目(ZD2020318);河北省教育厅青年基金项目(QN2023185)。

摘　　要：针对无人自动充电桩工作过程中背景复杂导致电动汽车充电口识别精度低等问题,提出基于改进YOLOv5算法的一种电动汽车充电口识别方法。采用融入加权双向特征金字塔结构增强不同层级间的信息融合能力;引入GhostNet网络结构中的深度可分离卷积GhostConv,替代原模型中的特征提取网络中的普通卷积层,减小了模型的计算开销;主干网络中使用SENet结构增加感受野信息,提高模型对充电口特征的提取能力;同时改进模型损失函数,引入EIoU损失函数代替原始CIoU损失函数,提高边界框回归精度。实验结果表明:改进后的模型在自制的多样化电动汽车充电口数据集上相较于原始YOLOv5算法在模型体积上减小了6.94 MB,检测精度提升到94.75%。同时与目前主流的检测算法对比,检测精度与检测速度也具有一定的优越性,适用于复杂背景环境下的电动汽车充电口的目标检测。To address the low accuracy in electric vehicle charging port recognition due to the complex background during the operation of unmanned automatic charging stations,this paper proposes a charging port recognition method for electric vehicles based on the improved YOLOv5 algorithm.First,the method incorporates a weighted bidirectional feature pyramid structure to enhance information fusion capabilities between different levels.Second,it introduces GhostConv,a depthwise separable convolution from the GhostNet network structure,replacing the ordinary convolution layers in the original feature extraction network,reducing the computational overhead of the model.The main network employs the SENet structure to increase the receptive field information,enhancing the model’s ability to extract charging port features.Meanwhile,the loss function of the model is improved by introducing the EIoU loss function to replace the original CIoU loss function,enhancing the accuracy of bounding box regression.Our experimental results demonstrate the improved model,compared to the original YOLOv5 algorithm,reduces the model size by 6.94 MB and achieves a detection accuracy of 94.75%on a self-made,diverse dataset of electric vehicle charging ports.Furthermore,compared to the mainstream detection algorithms,it delivers superior detection accuracy and speed,making it suitable for target detection of electric vehicle charging ports in complex background environments.

关 键 词：图像处理 目标检测 电动汽车充电口 注意力机制 多尺度融合 YOLOv5 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]