深度学习的接触网小目标缺陷识别研究  

Identification of the catenary small target defects in deep learning

作　　者：顾桂梅[1] 王小亮 Gu Guimei;Wang Xiaoiang(School of Automation and Electrical Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China)

机构地区：[1]兰州交通大学自动化与电气工程学院,兰州730070

出　　处：《电子测量与仪器学报》2024年第4期151-160,共10页Journal of Electronic Measurement and Instrumentation

基　　金：甘肃省科技计划资助项目(20JR10RA216)资助。

摘　　要：吊弦线夹螺栓是铁路接触网供电线路的重要器件,其状态会影响电力机车受流质量,于是对SSD算法进行改进:首先引入一种轻量级神经网络MobileNetV3用于前端特征提取,降低模型复杂度,以提高检测速度;其次采用CA注意力机制替换反向残差结构线性瓶颈层的SE模块,使位置信息沿空间两个方向聚合,调整后的特征层能够捕获全局远程特征信息;最后设计了特征融合模块以重构特征层,优化小目标检测层以提高对小目标的识别效果。还用CycleGAN等方法扩充训练样本,解决数据集不足的问题。实验结果表明,改进算法的模型复杂度下降,mAP@0.5和FPS分别达到95.5%和81 fps,该研究有助于接触网检测仪器向小型移动嵌入式设备转变。The dropper clamp bolt is an important component of railway power supply line,which can affect the flow quality of electric locomotive.Therefore,this paper improves the SSD algorithm:Firstly,a lightweight neural network MobileNetV3 is introduced for frontend feature extraction to reduce the model complexity and improve the detection speed;secondly,CA attention mechanism to replace the SE module of the linear bottleneck layer with inverted residuals structure,aggregate the position information in the two directions of space,and the adjusted feature layer can capture the global remote feature information.Finally,the feature fusion module for reconstructing the feature layer is designed to adjust the small target detection layer to improve the detection effect of small targets.This paper also expands the training sample with CycleGAN to solve the problem of insufficient data set.The experimental results show that the model complexity of the improved algorithm decreased,and mAP@0.5 and FPS reached 95.5% and 81 fps,respectively.This study helps the transformation of catenary detection instruments to small mobile embedded devices.

关 键 词：深度学习 小目标 注意力机制 接触网检测 

分 类 号：U225.4[交通运输工程—道路与铁道工程] TP391.4[自动化与计算机技术—计算机应用技术]