• Optics and Precision Engineering
  • Vol. 31, Issue 12, 1816 (2023)
Baoqing GUO1,2,* and Defen ZHANG1
Author Affiliations
  • 1School of Mechanical, Electronic and Control Engineering, Beijing Jiaotong University, Beijing 00044, China
  • 2Frontiers Science Center for Smart High-speed Railway System, Beijing Jiaotong University, Beijing 100044, China
  • show less
    DOI: 10.37188/OPE.20233112.1816 Cite this Article
    Baoqing GUO, Defen ZHANG. Railway few-shot intruding objects detection method with metric meta learning[J]. Optics and Precision Engineering, 2023, 31(12): 1816 Copy Citation Text show less
    Overall network framework
    Fig. 1. Overall network framework
    Structure of feature extraction network
    Fig. 2. Structure of feature extraction network
    Structure of residual block
    Fig. 3. Structure of residual block
    Improved channel attention module
    Fig. 4. Improved channel attention module
    Network pre-training strategy
    Fig. 5. Network pre-training strategy
    Thermal map effect of railway datasets
    Fig. 6. Thermal map effect of railway datasets
    Sample distribution with and without model pre-training and center related loss
    Fig. 7. Sample distribution with and without model pre-training and center related loss
    编号场景类别数量/张
    1空场景40
    2行人入侵40
    3泥石流入侵40
    4落石入侵40
    5列车经过40
    Table 1. Railway dataset composition

    算法 1:类中心微调(Pseudo Code of center fine tuning)

    输入:支持集S={(f(x1),y1),,(f(xN),yN)},其中yi1,...,kDk表示支持集S中所有标签为yi=k的样本集合

    输出:更新得到的每个中心值{c1,...,ck}

    1.初始化每个类中心值:

    for k in {1,...,k} do

      ck(φ)1NxiSkfϕk(xi)

       end for

    2.归一化支持集中样本xi到每个初始类中心的距离:

      for i in 1,...,N do

       p(φ)(y=k|xi)e-d(fφ(xi),ck(φ))k'e-d(fφ(xi),ck'(φ))

      end for

    3.计算损失并更新类中心:

        L(φ)-1Ni=1Nlog(p(φ)(y=k|xi))

    Table 1. [in Chinese]

    算法2:特征映射网络学习算法

    (Feature mapping network parameter learning algorithm)

    输入:支持集S={(f(x1),y1),,(f(xN),yN)},其中yi1,,k,查询集Q={(f(x1),y1),,(f(xm),ym)},特征映射网络初始化参数θC,超参数λα,迭代次数t0

    输出:特征映射网络参数θC

    1:while not converge do:

    2:  tt+1

    3:计算总的损失:

    Ltotal=Ls+Lc=-1mi=1mlog(e-d(fϕ(xi),ck)k'e-d(fϕ(x),ck'))+λi=1mxi-ck2

    4:  初始化类别中心:

       for i in k:

    ck=1Sk(xi,yi)Skf(xi)

       end for

    5:  计算反向传播误差:Lxi=Lsxi+λLcxi

    6:  更新参数θCθC=θC-αimLxixiθC

    7:end while

    Table 2. [in Chinese]
    5-way Acc
    Model1-shot5-shot
    MAM1748.7063.11

    RelationNet9

    ProtoNet8

    MatchNet18

    本文算法

    50.44

    48.51±0.40

    43.40±0.78

    50.65±0.45

    65.32

    67.09±0.36

    51.09±0.71

    74.40±0.33

    Table 2. Experiment results on MiniImageNet dataset(%)
    5-way Acc
    Model1-shot5-shot
    MAML60.23±0.2175.26±0.24

    RelationNet

    ProtoNet

    MatchNet

    本文算法

    61.95±0.32

    62.40±0.36

    58.45±0.23

    65.91±0.35

    76.68±0.40

    81.83±0.24

    65.23±0.35

    85.44±0.25

    Table 3. Experiment results on railway datasets
    目标类别准确率
    背景、泥石流、列车95.53±0.18
    背景、泥石流、列车、落石91.48±0.21
    背景、泥石流、列车、落石、行人85.44±0.25
    Table 4. mAp on different kinds of object
    网络结构5-way 1-shot5-way 5-shot
    基本网络59.01±0.4177.53±0.25
    基本网络+CBAM58.12±0.2675.85±0.23
    基本网络+改进CAM62.76±0.4178.39±0.24
    Table 5. Experiments on different attention mechanism(%)
    基本网络改进CAM类中心微调中心相关损失模型预训练5-way 1-shot5-way 5-shot
    59.01±0.4177.53±0.25
    62.76±0.4178.39±0.24
    -80.58±0.24
    64.90±0.3979.44±0.22
    64.16±0.4084.23+0.23
    65.91±0.3585.44±0.25
    Table 6. Ablation experiments on railway datasets
    N-way K-shotAccuracy
    5-way 5-shot85.44±0.25
    5-way 10-shot88.16±0.22
    5-way15-shot91.21±0.21
    5-way 20-shot93.51±0.21
    Table 7. Model accuracy under different K-shot
    Baoqing GUO, Defen ZHANG. Railway few-shot intruding objects detection method with metric meta learning[J]. Optics and Precision Engineering, 2023, 31(12): 1816
    Download Citation