Laser & Optoelectronics Progress, Volume. 60, Issue 6, 0610008(2023)

Improved ResNet Image Classification Model Based on Tensor Synthesis Attention

Yunfei Qiu1, Jiaxin Zhang1、*, Hai Lan2, and Jiaxu Zong3
Author Affiliations
  • 1College of Software, Liaoning Technical University, Huludao 125105, Liaoning, China
  • 2Quanzhou Institute of Equipment Manufacturing Haixi Institutes, Chinese Academy of Sciences, Quanzhou 362216, Fujian, China
  • 3Yuanqi Industrial Technology Company, Qingdao 266000, Shandong, China
  • show less
    Figures & Tables(19)
    Structure of residual block
    Structure of self-attention
    Structure of Synthesizer
    Overview of RTSA Net-101 model structure
    Three-tensor product
    Tensor synthesis attention module
    TSAR module
    Comparison of classification accuracy of different methods on Gaussian noise images
    Comparison of the classification accuracy of different methods for rotating images
    Comparison of the classification accuracy of different methods for the center cropped image
    • Table 1. Details of datasets

      View table

      Table 1. Details of datasets

      DatasetNumber of dataCategoryTraining setTesting set

      CIFAR-10

      CIFAR-100

      SVHN

      60000

      60000

      99289

      10

      100

      10

      50000

      50000

      73257

      10000

      10000

      26032

    • Table 2. Abbreviations of various methods

      View table

      Table 2. Abbreviations of various methods

      Method abbreviationConcrete model
      NoneCNN
      CACNN+Attention
      CSDCNN+Synthesizer Dense
      CFSDCNN+Factorized Synthesizer Dense
      CFSRCNN+Factorized Synthesizer Random
      CMSCNN+Mixture Synthesizers
      CTSA(CCNN+Tensor Synthetic Attention Channel
      CTSA(HCNN+Tensor Synthetic Attention Hight

      CTSA(W

      CTSAR

      CNN+Tensor Synthetic Attention Width

      CNN+Tensor Synthetic Attention Random

    • Table 3. Classification accuracy results of different methods for Gaussian noise images

      View table

      Table 3. Classification accuracy results of different methods for Gaussian noise images

      MethodGaussian noise
      0.010.020.030.040.050.060.070.080.090.10
      None45.9634.3527.8022.8320.0117.8616.2315.0714.4513.76
      CA47.0235.8329.2124.7021.4919.3317.4116.5115.4214.80
      CSD48.7137.7931.5427.3023.7022.2320.1818.9217.4517.03
      CFSR46.5137.2831.6927.5624.1422.6020.6219.7518.2517.76
      CFSD48.9037.5030.4225.8922.8520.7019.3317.9116.8716.07
      CMS48.2036.8029.5525.1321.4519.3417.7416.6115.7615.26
      CTSA(C49.2936.8131.9828.1125.4923.7522.3921.4420.0919.54
      CTSA(H44.6630.9223.2619.5817.0815.1214.0413.3512.6112.26

      CTSA(W

      CTSAR

      42.20

      45.93

      30.21

      34.02

      23.65

      27.06

      19.92

      23.01

      17.53

      19.91

      16.41

      17.36

      15.18

      16.02

      14.65

      14.93

      14.33

      13.93

      13.64

      13.40

    • Table 4. Classification accuracy results of different methods for rotating pictures

      View table

      Table 4. Classification accuracy results of different methods for rotating pictures

      MethodRotation angle/(°)
      306090120150180210240270300330
      None37.5826.5123.0721.1219.4019.3918.6818.8918.2717.7417.40
      CA36.5436.5723.2721.2519.7519.4419.7018.6118.1918.4617.75
      CSD37.9038.1723.2421.7919.2619.4519.3618.5117.5517.6017.98
      CFSR37.0426.8623.1020.6919.7218.6918.7018.2717.6917.2417.08
      CFSD39.2728.8523.5522.8520.3719.8519.9518.7118.8618.3417.88
      CMS37.9126.9322.9621.7419.6118.8918.7618.0118.0617.6217.65
      CTSA(C38.7138.3624.5121.8020.9720.3620.8919.9319.8619.1220.00
      CTSA(H42.3230.4426.2824.5322.1722.1822.3721.8320.8520.0720.16

      CTSA(W

      CTSAR

      45.53

      33.72

      38.64

      24.70

      27.21

      22.06

      25.20

      19.88

      22.39

      19.00

      22.23

      18.17

      22.03

      18.03

      21.27

      17.22

      20.61

      17.03

      20.01

      17.23

      19.89

      16.79

    • Table 5. Classification accuracy results of different methods for center cropped images

      View table

      Table 5. Classification accuracy results of different methods for center cropped images

      MethodCenter crop size/pixel
      1015202530
      None14.8920.2232.6851.1561.20
      CA13.2017.9828.0047.4160.58
      CSD13.7919.2930.1548.8561.38
      CFSR14.0319.1930.2348.8961.23
      CFSD13.5319.0331.2851.8763.26
      CMS15.1220.1232.5851.3162.16
      CTSA(C13.7418.9332.6053.5663.62
      CTSA(H14.7020.1531.5150.5661.84

      CTSA(W

      CTSAR

      15.22

      14.03

      20.53

      19.19

      33.23

      30.23

      52.12

      48.89

      62.21

      61.23

    • Table 6. Classification results of different methods

      View table

      Table 6. Classification results of different methods

      MethodAccuracy /%Recall /%Precision /%F1 /%
      None83.5262.3152.3071.59
      CA82.3071.2868.4682.37
      CSD84.2380.5685.7778.38
      CFSR72.9473.5983.2180.97
      CFSD87.7181.3277.3267.98
      CMS65.2388.5760.5856.35
      CTSA(C91.6590.2089.9186.72
      CTSA(H90.5089.3086.3584.23
      CTSA(W89.9687.2387.2583.37
      CTSAR88.6788.5989.3482.56
    • Table 7. Comparison of classification results of residual network models

      View table

      Table 7. Comparison of classification results of residual network models

      ModelAccuracy /%
      ResNet-1889.33
      ResNet-3488.76
      ResNet-5086.72
      ResNet-10193.12
    • Table 8. Comparison of classification accuracy of different RTSA Net models

      View table

      Table 8. Comparison of classification accuracy of different RTSA Net models

      ModelAccuracy /%
      RTSA Net-1890.30
      RTSA Net-3491.23
      RTSA Net-5092.98
      RTSA Net-10196.12
    • Table 9. Comparison with other advanced methods for classification accuracy and average test running time

      View table

      Table 9. Comparison with other advanced methods for classification accuracy and average test running time

      ModelCIFAR-10CIFAR-100SVHN
      Accuracy /%running time /sAccuracy /%running time /sAccuracy /%running time /s
      ANODE2560.600.037883.500.0308
      Sign-symmetry2680.980.034152.250.030189.840.0317
      Improved GAN2788.170.029191.890.0269
      CLS-GAN2891.700.043194.020.0277
      NNCLR2993.700.028979.000.0289
      CCT-6/3x13095.290.032177.310.0271
      ResNet56 with reSGHMC3196.100.027880.140.0278
      RTSA Net-10196.120.025881.600.026096.670.0262
    Tools

    Get Citation

    Copy Citation Text

    Yunfei Qiu, Jiaxin Zhang, Hai Lan, Jiaxu Zong. Improved ResNet Image Classification Model Based on Tensor Synthesis Attention[J]. Laser & Optoelectronics Progress, 2023, 60(6): 0610008

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category: Image Processing

    Received: Oct. 29, 2021

    Accepted: Jan. 17, 2022

    Published Online: Mar. 7, 2023

    The Author Email: Jiaxin Zhang (1595775491@qq.com)

    DOI:10.3788/LOP212836

    Topics