基于多尺度可分离蒸馏网络的列车轮对踏面缺陷检测算法  

Train wheel tread defect detection algorithm based on multi-scale separable distillation network

作　　者：程翔 朱禹熹 贾林[3] CHENG Xiang;ZHU Yuxi;JIA Lin(School of Artificial Intelligence,Hunan Railway Professional College,Zhuzhou 412007,China;School of Materials Science and Engineering,Chongqing University of Technology,Chongqing 400054,China;School of Rail Transit,Hunan University of Technology,Zhuzhou 412001,China)

机构地区：[1]湖南铁道职业技术学院人工智能学院,湖南株洲412007 [2]重庆理工大学材料科学与工程学院,重庆400054 [3]湖南工业大学轨道交通学院,湖南株洲412001

出　　处：《铁道科学与工程学报》2024年第11期4789-4803,共15页Journal of Railway Science and Engineering

基　　金：国家自然科学基金资助项目(62303178);湖南省自然科学基金资助项目(2023JJ60232);湖南省教育厅资助科研项目(23B1018,22B0577)。

摘　　要：作为列车安全维护的重点问题,列车轮对表面缺陷检测性能易受低算力、多尺度和复杂背景等因素的制约。这使得许多基于深度学习的目标检测算法难以完全发挥其性能。针对以上问题,提出一种基于多尺度可分离蒸馏网络的列车轮对踏面缺陷检测算法(D-MSCNet)。在该方法中,首先设计一种由跨层可分离特征提取模块(CSEM)和多尺度可分离下采样模块(MSDM)组成的主干网络(MSA-ResNet)。其中,CSEM模块通过在残差架构中引入深度可分离卷积、扩展卷积、通道压缩和跨层连接机制来保证它能在不显著增加计算量的同时丰富特征信息、扩大感受野。多尺度可分离下采样模块(MSDM)通过引入多尺度特征融合机制和注意力模块以在下采样任务过程中提高多尺度特征信息、弱化背景信息。其次,设计了一种新的主从区域知识蒸馏策略来有效地压缩和简化模型。它首先根据标签中的Ground Truth划分主要和次要蒸馏区域,然后将两者作用于各个回归分支之间以执行知识蒸馏任务。最后,在实际列车轮对踏面数据集上进行了实验分析,对比实验表明当所提方法D-MSCNet在加载小参数网络(MSA-ResNet18)时,其M_(AP)=64.9%、F_(PS)=85优于大多数对比方法,表明了该方法能够有效地平衡检测速度和检测精度。此外,通过消融实验、模块对比实验以及可视化分析进一步验证了所提方法的有效性和优越性。As an important aspect of train safety maintenance,the performance of surface defect detection on train wheelsets is constrained by factors such as low computing power,multi-scale,and complex backgrounds.This makes it difficult for many classic object detection algorithms based on deep learning to fully exploit their performance.According to the above problems,this paper proposed a train wheel tread defect detection algorithm based on a multi-scale separable distillation network,called D-MSCNet.According to the method,a backbone network(MSA-ResNet)consisting of a cross layer separable feature extraction module(CSEM)and a multi-scale separable down-sampling module(MSDM)was designed in this paper.Among them,the depth-wise separable convolution,dilated convolution,channel compression structure,and cross-layer connection mechanism were designed in the CSEM module to enrich feature information and expand the receptive field without significantly increasing the amount of computation.The multi-scale separable down-sampling module(MSDM)extracted multi-scale feature information and weakened background information during down-sampling tasks by introducing multi-scale feature fusion mechanisms and attention modules.Then,a new master-slave region knowledge distillation strategy was designed to effectively compress and simplify the detection model.It first divided the primary and secondary distillation regions according to the ground truth information in the label,and then applied the two between each regression branch to perform the knowledge distillation task.Finally,experimental analysis was conducted on an actual train wheel-tread dataset.Comparative experiments show that when the proposed method D-MSCNet is loaded with a small parameter network(MSA-ResNet18),its MAP=64.9%and FPS=85 are better than most of comparative methods,indicating that this method can effectively balance detection speed and detection accuracy.In addition,the effectiveness and superiority of the proposed method are further verified through ablation e

关 键 词：轮对踏面缺陷 知识蒸馏 无锚检测器 多尺度特征模块 深度学习 

分 类 号：U279[机械工程—车辆工程]