基于轻量化GoogLeNet模型的轨道扣件缺陷状态识别  

Rail Fastener Defect Status Identification Based on Lightweight GoogLeNet Model

作　　者：李少佳 胡美振 陈辉东[1] 刘艳霞[1] LI Shaojia;HU Meizhen;CHEN Huidong;LIU Yanxia(College of Urban Rail Transit and Logistics,Beijing Union University,Beijing 100101,China)

机构地区：[1]北京联合大学城市轨道交通与物流学院,北京100101

出　　处：《北京联合大学学报》2023年第1期6-12,共7页Journal of Beijing Union University

基　　金：北京联合大学人才强校优选—拔尖计划项目(BPHR2020BZ02);北京市教委科技计划项目(KM202111417004,KM202011417004,KM201911417007);北京联合大学科研项目(ZK30202002);北京联合大学教育教学研究与改革重点项目(JY2021Z002)。

摘　　要：我国是交通大国,并正在向交通强国迈进。轨道维护至关重要,其中扣件的健康状态对于列车的运行安全不容忽视。然而,由于经典卷积神经网络模型的复杂度较高,尤其在识别速度方面无法满足轨道扣件状态识别任务对实时性的要求。鉴于此,设计了一种轻量化GoogLeNet网络模型,在保障模型精度的同时提升模型的推理速度。实验结果表明,轻量化GoogLeNet网络模型的分类精度为92.7%,FPS达到了254.2。相比于VGG16、VGG19和原始的GoogLeNet模型,其识别精度分别提高了21.7、19.5和8.6个百分点,单张图片的推理速度分别减少了3.637 ms、4.8247 ms和2.9432 ms。China is a big transportation nation and is developing into a powerful transportation country.For the safety of train operation,track maintenance is extremely vital,and the fasteners sound condition cannot be disregarded.However,the standard convolutional neural network model cannot satisfy the real-time demands of the track fastener status recognition job due to its high level of complexity,particularly in the recognition speed.In view of this,a lightweight GoogLeNet network model is designed to improve the inference speed of the model while maintaining the model s accuracy.According to the testing findings,the classification accuracy of the lightweight GoogLeNet network model is 92.7%and the FPS reaches 254.2.Compared with VGG16,VGG19 and the original GoogLeNet model,the recognition accuracy is increased by 21.7,19.5 and 8.6 percent point,respectively.The inference speed of a single image is reduced by 3.637 ms,4.8247 ms and 2.9432 ms,respectively.

关 键 词：轻量化 GoogLeNet 图像增强 扣件状态分类 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]