基于改进网络通道剪枝算法的接触网零部件实时检测  

Real-time Detection of Overhead Contact Line Components Based on Improved Network Channel Pruning Algorithm

作　　者：宋超 于龙[1,2] 张冬凯 单禹 王健 杨珂浩 Song Chao;Yu Long;Zhang Dongkai;Shan Yu;Wang Jian;Yang Kehao(Key Laboratory of Railway Industry on Smart Traction Power Supply,Southwest Jiaotong University,Sichuan,Chengdu 611756,China;School of Electrical Engineering,Southwest Jiaotong University,Sichuan,Chengdu 611730,China;Research Institute of Tangshan,Southwest Jiaotong University,Hebei,Tangshan 063000,China;School of Electric Power Engineering,Kunming University of Science and Technology,Yunnan,Kunming 650031,China)

机构地区：[1]西南交通大学智能牵引供电铁路行业重点实验室,四川成都611730 [2]西南交通大学电气工程学院,四川成都611730 [3]西南交通大学唐山研究院,河北唐山063000 [4]昆明理工大学电力工程学院,云南昆明650031

出　　处：《铁道技术标准（中英文）》2023年第8期30-39,共10页Railway Technical Standard(Chinese & English)

基　　金：四川省自然科学基金(2022NSFSC0572)。

摘　　要：为满足接触网零部件在线检测的要求,提出一种改进的网络通道剪枝算法,极大提升了利用深度学习方法进行接触网零部件检测的实时性。以YOLOv5s作为零部件检测的深度网络,通过在损失函数中加入L1正则项来对网络进行稀疏训练,以得到一个稀疏模型。引入尺度缩放因子来评估稀疏模型中通道的重要性,裁剪具有较小缩放因子的通道和滤波器。在此基础上,提出一种Bottleneck残差结构的修剪方法,进一步提高网络的压缩比。同时,构建一种网络通道规整化策略,使剪枝后网络每层保留的通道数为8的倍数,从而保持原网络的规整性,优化网络在硬件设备上的推理速度。实验结果表明,该方法对10类零部件的mAP值为83.7%,仅比YOLOv5s低0.2%,但模型的参数量减少了93%,模型存储空间压缩到原来的10%。此外,检测速度从85 FPS提高到154 FPS,提升了81%,基本满足高速铁路接触网零部件的实时检测需求。To meet the requirements of online detection of overhead contact line(OCL)components,an improved network channel pruning algorithm is proposed,significantly enhancing the real-time performance of utilizing deep learning methods for OCL component detection.In this paper,YOLOv5s is employed as the deep network for component detection.Firstly,L1 regular constraints are incorporated into the loss function of YOLOv5s network to perform sparse training on the network,so as to obtain a sparse model for pruning.Next,a scaling factor is introduced to assess the importance of each channel in the sparse model,and channels with smaller scaling factors and their corresponding filters are pruned.Furthermore,a Bottleneck residual structure pruning method is proposed to further increase the compression ratio of the network.Simultaneously,a network channel regularization strategy is devised to ensure that the number of channels retained by each layer of the pruned network is a multiple of 8,thereby maintaining the regularity of the original network and optimizing the inference speed of the network on hardware devices.Experimental results demonstrate that the proposed method achieves an mAP value of 83.7%for 10 types of OCL components,which is only 0.2%lower than YOLOv5s.However,the model parameters are reduced by 93%,and the model storage space is compressed to 1/10 of the original size.In addition,the detection speed has increased from 85 FPS to 154 FPS,an improvement of 81%,effectively meeting the real-time detection requirements of OCL components for high-speed railways.

关 键 词：接触网 深度学习 稀疏训练 通道剪枝 实时检测 

分 类 号：U225.4[交通运输工程—道路与铁道工程]