基于改进YOLOv8s的铁路车站信号平面布置图信息提取方法  

Information Extraction Method of Railway Station Signal Plan Layout Based on Improved YOLOv8s

作　　者：郑云水[1] 蒙阳 ZHENG Yunshui;MENG Yang(School of Automation and Electrical Engineering,Lanzhou Jiaotong University,Lanzhou Gansu 730070,China)

机构地区：[1]兰州交通大学自动化与电气工程学院,甘肃兰州730070

出　　处：《中国铁道科学》2024年第5期209-220,共12页China Railway Science

基　　金：国家重点研发计划项目(2021YFB2300305-05)。

摘　　要：针对车站信号平面布置图在2次利用过程中,无法将不同设备图元转化为通用数据的问题,提出基于改进YOLOv8s的信息提取方法。首先,将空间和通道重建卷积(SCConv)与颈部网络中的C2f单元相结合,以降低特征图的冗余度,减少空间和通道维度上的重复信息,实现检测模型的轻量化;其次,在传统YOLOv8s模型的主干和颈部网络中引入高效多尺度注意力模块(EMA),以增强模型对图内设备周边上下文信息的获取能力;最后,将模型中的标准卷积层替换为具有更大感受野的注意力卷积(RFAConv),以解决卷积核参数共享所导致的特征提取限制。结果表明:所提方法的平均精度均值可达92.5%,相较于YOLOv8s算法提升6.6%;综合评价指标值达到91.7,且模型权重参数减少16%,优于大多数常规模型。该方法不仅能够高效地从图纸中提取信号设备的布局信息,还为自动生成联锁数据配置文件和提高联锁软件搭建效率提供了有效的技术手段。To address the issue that various equipment primitives cannot be converted into common data during the secondary use of railway station signal plan layout diagrams,an information extraction method based on improved YOLOv8s is proposed.Firstly,the spatial and channel reconstruction convolutional block(SCConv)is combined with the C2f unit in the neck network to reduce the redundancy of feature maps and repetitive information across spatial and channel dimensions,so as to facilitate a lightweight detection model.Secondly,the efficient multi-scale attention module(EMA)is introduced into the backbone and neck network of the traditional YOLOv8s model to enhance the model’s acquisition capability,obtaining contextual information around the equipment in the diagram.Finally,the standard convolution layer in the model is replaced by the attention convolution(RFAConv)with a larger receptive field to solve the feature extraction limitation stem from the sharing of convolution kernel parameters.The results show that the proposed method achieves an average precision of 0.925,a 6.6%higher improvement over the YOLOv8s algorithm;its comprehensive evaluation index value reaches 0.917,and the model weight parameters are reduced by 16%,outperforming most standard models.This method efficiently extracts the layout information of signal equipment from drawings,providing a potent technical solution for automatically generating interlocking data configuration files and improving the efficiency of interlocking software construction.

关 键 词：信号平面布置图 图像识别 注意力模块 信息提取 YOLOv8s 

分 类 号：U284.39[交通运输工程—交通信息工程及控制]