• Laser & Optoelectronics Progress
  • Vol. 60, Issue 6, 0610005 (2023)
Wenliang Wang, Xiaodi Yang*, Boya Zhang, Jishun Ma, Peng Zeng, and Peng Han
Author Affiliations
  • CSSC (Zhejiang) Ocean Technology Co., Ltd., Zhoushan 316000, Zhejiang, China
  • show less
    DOI: 10.3788/LOP213033 Cite this Article Set citation alerts
    Wenliang Wang, Xiaodi Yang, Boya Zhang, Jishun Ma, Peng Zeng, Peng Han. Application of a Lightweight Convolutional Neural Network in Ship Classification[J]. Laser & Optoelectronics Progress, 2023, 60(6): 0610005 Copy Citation Text show less
    Ghost module
    Fig. 1. Ghost module
    Ghost bottleneck module. (a) Bottleneck module with a step size of 1; (b) bottleneck module with a step size of 2
    Fig. 2. Ghost bottleneck module. (a) Bottleneck module with a step size of 1; (b) bottleneck module with a step size of 2
    ACNet module
    Fig. 3. ACNet module
    Asymmetric Ghost module
    Fig. 4. Asymmetric Ghost module
    Asymmetric Ghost bottleneck module. (a) Bottleneck module with a step size of 1; (b) bottleneck module with a step size of 2
    Fig. 5. Asymmetric Ghost bottleneck module. (a) Bottleneck module with a step size of 1; (b) bottleneck module with a step size of 2
    Comparison of the overall structure of each network
    Fig. 6. Comparison of the overall structure of each network
    Comparison curve of test set precision
    Fig. 7. Comparison curve of test set precision
    Comparison curve of test set loss value
    Fig. 8. Comparison curve of test set loss value
    InputOperatorExpOutSEStride
    2242×3Conv 2d 3×3-8-2
    1122×8AG-bneck88-1
    1122×8AG-bneck2412-2
    562×12AG-bneck3612-1
    562×12AG-bneck362012
    282×20AG-bneck602011
    282×20AG-bneck12040-2
    142×40AG-bneck10040-1
    142×40AG-bneck9240-1
    142×40AG-bneck9240-1
    142×40AG-bneck2405611
    142×56AG-bneck3365611
    142×56AG-bneck3368012
    72×80AG-bneck48080-1
    72×80AG-bneck4808011
    72×80AG-bneck48080-1
    72×80AG-bneck4808011
    72×80Conv 2d 1×1-480-1
    72×480AvgPool 7×7----
    12×480FC-classes--
    Table 1. Detail parameters of AGNet
    IDClassTrain_setTest_set
    15470104
    2332938636
    3594408986
    Table 2. Information table of sample set
    ModelMultiply-accumulate operations /106Parameters /106Accuracy /%Speed /(frame·s-1
    AGNet49.350.7293.8747.76
    AGNet-large49.971.3692.5041.79
    GhostNet-5045.841.3590.9443.61
    GhostNet-50-small45.220.7192.0350.13
    Table 3. Comparison of each network evaluation indicators
    ModelMultiply-accumulate operations /106Parameters /106Accuracy /%Speed /(frame·s-1
    AGNet49.350.7393.5546.93
    AGNet-large49.971.3990.3241.24
    GhostNet-5045.841.3887.6042.97
    GhostNet-50-small45.220.7292.9449.64
    Table 4. Comparison of each network performance
    ModelMultiply-accumulate operations /106Parameters /106Accuracy /%
    MobileNetv3-small-1002470.062.5487.81
    ResNet1811821.6611.6989.84
    DPN-68b262338.5512.6191.87
    RegNetx-00227200.692.6889.53
    GhostNet-502345.841.3592.50
    AGNet49.350.7293.87
    Table 5. Comparison results of different classification networks
    Wenliang Wang, Xiaodi Yang, Boya Zhang, Jishun Ma, Peng Zeng, Peng Han. Application of a Lightweight Convolutional Neural Network in Ship Classification[J]. Laser & Optoelectronics Progress, 2023, 60(6): 0610005
    Download Citation