Acta Optica Sinica, Volume. 45, Issue 9, 0928001(2025)

SAR Ship Detection Method Based on Ladder Residual and Coordinate Information Recombination

Wenxing Liu1, Huilin Shan2、*, Xingtao Wang1, Jieru Liu3, Ge Chen1, and Mengjiao Shan1
Author Affiliations
  • 1School of Electronic and Information Engineering, Nanjing University of Information Science and Technology, Nanjing 210044, Jiangsu , China
  • 2Wuxi University, Wuxi 214000, Jiangsu , China
  • 3South China Normal University, Guangzhou 510631, Guangdong
  • show less
    Figures & Tables(22)
    SRCDNet structure diagram
    Diagrams of conventional convolution, deep separable convolution, and partial convolution modules
    SRP diagram
    CBAM diagram
    Diagram of SCAA-Net module
    CIRConv module diagram
    Fn module diagram
    CIRConv feature map visualization
    Visualization comparison of image features before and after outputting. (a) Before outputting image; (b) y direction coordinate recombination image; (c) x direction coordinate recombination image; (d) combined output image
    Time frequency waveform analysis of CIRConv mentioned in this article
    Time frequency 3D surface maps of CIRConv. (a) Small ship; (b) large ship; (c) moving ship
    Comparison of real detection effects between Ori and SRCDNet. (a) Real values; (b) detection of Ori model; (c) SRCDNet detection
    Comparison of algorithm detection results. (a1)‒(a5) True values; (b1)‒(b5) Center Net detection; (c1)‒(c5) Faster R-CNN detection; (d1)‒(d5) YOLOv8 detection; (e1)‒(e5) YOLOv7 detection; (f1)‒(f5) SRCDNet detection
    Comparison of losses of SRCDNet and original model
    Comparison of mAP0.5 between SRCDNet and original model
    Comparison of mAP0.5∶0.95 between SRCDNet and original model
    • Table 1. Parameter configuration of hardware and software platforms

      View table

      Table 1. Parameter configuration of hardware and software platforms

      ConfigurationModel
      CPU13rd Gen Intel Core i7-13700KF
      GPUNVIDIA GeForce RTX 4080
      Operating systemWindows 10
      Running memory32 GB
      Video memory16 GB
      Language and frameworkAnconda3+Python3.8+Pytorch1.13.0
    • Table 2. Training parameter configuration

      View table

      Table 2. Training parameter configuration

      Training parameterValue
      Momentum parameter0.937
      Initial learning rate0.001
      Batch size8
      Epochs300
      OptimizerSGD
    • Table 3. Comparison of improvement effects in ablation experiments

      View table

      Table 3. Comparison of improvement effects in ablation experiments

      ModelDatasetPRmAP0.5mAP0.5∶0.95FLOPs/109
      OriHRSID0.8760.7860.8630.602105.1
      Ori+SRPHRSID0.9150.7830.8740.61885.7
      Ori+SCAA-NetHRSID0.9000.7960.8810.614101.1
      Ori+SCAA-Net(Leaky ReLU)HRSID0.8920.7980.8770.608101.1
      Ori+CIRCovHRSID0.8970.7880.8740.615107.1
      Ori+SRP+SCAA-Net(Leaky ReLU)HRSID0.9170.8090.8900.62781.6
      Ori+SRP+SCAA-NetHRSID0.9010.7820.8750.61181.6
      Ori+SRP+SCAA-Net(Leaky ReLU)+CIRConvHRSID0.9190.8120.8930.62883.6
      Ori+SRP+SCAA-Net(Leaky ReLU)+CIRConv+NWDHRSID0.9240.8230.9040.62983.6
      OriSSDD0.8980.9370.9570.564105.1
      Ori+SRPSSDD0.9350.9350.9710.59185.7
      Ori+SCAA-NetSSDD0.9250.8960.9590.557101.1
      Ori+CIRCovSSDD0.9240.9290.9640.569107.1
      Ori+SCAA-Net(Leaky ReLU)SSDD0.9330.9390.9690.573101.1
      Ori+SRP+SCAA-Net(Leaky ReLU)SSDD0.9460.9110.9650.57681.6
      Ori+SRP+SCAA-NetSSDD0.9210.9350.9650.57281.6
      Ori+SRP+SCAA-Net(Leaky ReLU)+CIRConvSSDD0.9460.9450.9710.59083.6
      Ori+SRP+SCAA-Net(Leaky ReLU)+CIRConv+NWDSSDD0.9470.9500.9850.59283.6
    • Table 4. Ablation experiments of backbone network

      View table

      Table 4. Ablation experiments of backbone network

      BackboneDatasetPRmAP0.5mAP0.5∶0.95FLOPs /109
      Polymeric structureHRSID0.8760.7860.8630.602105.1
      Conventional residual structureHRSID0.8540.7700.8420.58978.9
      Stepwise residual structureHRSID0.8820.7860.8590.591130.5
      Stepwise residual structure+PConvHRSID0.8870.7840.8630.59879.4
      Stepwise residual structure+DWConv, k=7HRSID0.9160.7960.8720.610103.5
      Stepwise residual structure+DWConv, k=5HRSID0.9120.7920.8700.59993.2
      Stepwise residual structure+DWConv, k=3HRSID0.9140.7890.8730.61285.7
      SRPHRSID0.9150.7830.8740.61885.7
      Polymeric structureSSDD0.8980.9370.9570.564105.1
      Conventional residual structureSSDD0.8800.9120.9320.55878.9
      Stepwise residual structureSSDD0.9140.9280.9630.567130.5
      Stepwise residual structure+PConvSSDD0.9180.9240.9650.56879.4
      Stepwise residual structure+DWConv, k=7SSDD0.9300.9320.9690.574103.5
      Stepwise residual structure+DWConv, k=5SSDD0.9240.9280.9650.57293.2
      Stepwise residual structure+DWConv, k=3SSDD0.9340.9310.9680.57785.7
      SRPSSDD0.9350.9350.9710.59185.7
    • Table 5. Ablation experiments of feature fusion parts

      View table

      Table 5. Ablation experiments of feature fusion parts

      NeckDatasetPRmAP0.5mAP0.5∶0.95
      Polymeric structureHRSID0.8760.7860.8630.602
      Polymeric structure +CAHRSID0.8750.7890.8610.599
      Polymeric structure +SAHRSID0.8720.7910.8650.604
      Polymeric structure +CA+SAHRSID0.8860.7940.8720.608
      SCAA-NetHRSID0.9000.7960.8810.614
      SCAA-Net (Leaky ReLU)HRSID0.8920.7980.8770.608
      Polymeric structureSSDD0.8980.9370.9570.564
      Polymeric structure +CASSDD0.8920.9330.9540.569
      Polymeric structure +SASSDD0.8940.9260.9560.566
      Polymeric structure +CA+SASSDD0.9130.9150.9580.559
      SCAA-NetSSDD0.9250.8960.9590.557
      SCAA-Net (Leaky ReLU)SSDD0.9330.9390.9690.573
    • Table 6. Comparison of proposed model and mainstream models

      View table

      Table 6. Comparison of proposed model and mainstream models

      ModelPRmAP
      Center Net91.270.6287.43
      SSD80.183.3482.01
      Faster R-CNN78.484.1080.60
      YOLOv889.385.2087.50
      Ori87.678.6086.30
      SRCDNet92.482.3090.40
    Tools

    Get Citation

    Copy Citation Text

    Wenxing Liu, Huilin Shan, Xingtao Wang, Jieru Liu, Ge Chen, Mengjiao Shan. SAR Ship Detection Method Based on Ladder Residual and Coordinate Information Recombination[J]. Acta Optica Sinica, 2025, 45(9): 0928001

    Download Citation

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

    Category: Remote Sensing and Sensors

    Received: Nov. 11, 2024

    Accepted: Feb. 25, 2025

    Published Online: May. 19, 2025

    The Author Email: Huilin Shan (shanhuilin@nuist.edu.cn)

    DOI:10.3788/AOS241731

    CSTR:32393.14.AOS241731

    Topics