• Optics and Precision Engineering
  • Vol. 31, Issue 17, 2598 (2023)
Yingxu LIU1, Chunyu PU1, Diankun XU2, Yichuan YANG1, and Hong HUANG1,*
Author Affiliations
  • 1Key Laboratory of Optoelectronic Technology and Systems of the Education Ministry of China, Chongqing University, Chongqing400044, China
  • 2Measurement and Control Technology and Instrument major, College of Optoelectronic Engineering, Chongqing University, Chongqing400044, China
  • show less
    DOI: 10.37188/OPE.20233117.2598 Cite this Article
    Yingxu LIU, Chunyu PU, Diankun XU, Yichuan YANG, Hong HUANG. Lightweight deep global-local knowledge distillation network for hyperspectral image scene classification[J]. Optics and Precision Engineering, 2023, 31(17): 2598 Copy Citation Text show less
    Framework of proposed Lightweight Deep Globle-local Knowledge Distillation(LDGLKD) network
    Fig. 1. Framework of proposed Lightweight Deep Globle-local Knowledge Distillation(LDGLKD) network
    Detail structure of layer l in ViT model
    Fig. 2. Detail structure of layer l in ViT model
    Overall build process for OHID-SC dataset
    Fig. 3. Overall build process for OHID-SC dataset
    Examples of scene in constructed OHID-SC dataset
    Fig. 4. Examples of scene in constructed OHID-SC dataset
    Examples and numbers of scene in HSRS-SC dataset
    Fig. 5. Examples and numbers of scene in HSRS-SC dataset
    OAs with respect to different temperatures
    Fig. 6. OAs with respect to different temperatures
    Effect of KD coefficient α on OAs
    Fig. 7. Effect of KD coefficient α on OAs
    Ablation experiment results
    Fig. 8. Ablation experiment results
    Confusion matrix results of LDGLKD network
    Fig. 9. Confusion matrix results of LDGLKD network
    OAs with different training data percentages for comparison methods
    Fig. 10. OAs with different training data percentages for comparison methods
    算法:轻量化深度全局-局部知识蒸馏
    训练过程:

    阶段1:以小的知识蒸馏系数α进行训练

    输入:训练集图像Iin,温度T,知识蒸馏系数α,教师模型ViT,学生模型VGG16,教师模型函数集合ftea(),学生模型函数集合fstu()

    输出:经过优化的教师模型与学生模型;

    1:Fteaftea(Iin)

    2:Fstufstu(Iin)

    3:LKDT2×KLdiv(QstuT,QteaT)QstuTQteaT公式(18)和(19)计算;

    4:LfirαLKD+(1-α)(LCtea+LCstu)

    5:以Lfir为损失函数对教师模型以及学生模型进行共同优化;

    阶段2:以大的知识蒸馏系数α进行训练
    输入:训练集图像Iin,温度T,知识蒸馏系数α,教师模型ViT,学生模型VGG16,教师模型函数集合ftea(),学生模型函数集合fstu()

    输出:经过优化的学生模型;

    6:Fteaftea(Iin)

    7:Fstufstu(Iin)

    8:LKDT2×KLdiv(QstuT,QteaT)QstuTQteaT公式(18)和(19)计算;

    9:Lsec2αLKD+(1-2α)LCstu

    10:以Isec为损失函数对学生模型进行优化,并停止对教师模型的训练;

    测试阶段:

    场景分类

    输入:测试集图像,Iintest,学生模型VGG16,学生模型函数集合,fstu()

    输出:预测标签Ylabel

    11.Fstufstu(Iintest)

    12.Ylabelsoftmax(Fstu)

    13.生成每一张测试集图像的语义标签Ylabel
    Table 1. Steps of LDGLKD algorithm
    MethodOHID-SCHSRS-SC
    ResNet10175.86±1.4195.95±0.26
    ResNet1880.34±0.3696.43±0.47
    GoogleNet58.17±3.8087.81±0.04
    EffcientNet83.81±0.6697.21±0.64
    VGG1681.78±0.5095.13±0.22
    SKAL-R55.14±0.4296.08±0.15
    SKAL-V59.80±0.4396.01±0.24
    ACRNet-R60.39±0.1895.75±0.07
    ACRNet-M65.02±0.1497.96±0.41
    LDGLKD91.62±0.2097.96±0.04
    Table 2. Results of comparison algorithm [OA±STD]
    MethodDatasetOHID-SCHSRS-SC
    ResNet101Train952
    Test45126
    ResNet18Train849
    Test19121
    GoogleNetTrain950
    Test27123
    EfficientNetTrain1051
    Test56125
    VGG16Train752
    Test15129
    SKAL-RTrain511
    Test1434
    SKAL-VTrain515
    Test1536
    ACRNet-RTrain24
    Test513
    ACRNet-MTrain25
    Test514
    LDGLKDTrain843
    Test13108
    Table 3. Running time of comparison algorithms
    Yingxu LIU, Chunyu PU, Diankun XU, Yichuan YANG, Hong HUANG. Lightweight deep global-local knowledge distillation network for hyperspectral image scene classification[J]. Optics and Precision Engineering, 2023, 31(17): 2598
    Download Citation