Optics and Precision Engineering, Volume. 31, Issue 6, 892(2023)

Ship detection for complex scene images of space optical remote sensing

Xinwei LIU1...2,3, Yongjie PIAO1,3,*, Liangliang ZHENG1,3, Wei XU1,3, and Haolin JI1,23 |Show fewer author(s)
Author Affiliations
  • 1Changchun Institute of Optics, Fine Mechanics and Physice,Chinese Academy of Sciences, Changchun30033, China
  • 2University of Chinese Academy of Sciences, Beijing100039, China
  • 3Key Laboratory of Space-Based Dynamic & Rapid Optical Imaging Technology, Chinese Academy of Sciences, Changchun100, China
  • show less
    Figures & Tables(18)
    Structure chart of IM-YOLO-s
    Schematic diagram of Focus operation process
    Structure chart of CoordinateAttention
    Network structure comparison diagram of Neck
    Network structure chart of CSPLayer
    Intersection of two boxes with different values IoU
    Schematic diagram of penalty item of central point of CIoU
    Training loss function diagram
    Comparison chart of actual detection effect of different algorithm
    Real target box annotation renderings
    Test results of YOLOX-s
    Test results of IM-YOLO-s
    • Table 1. Data set division

      View table
      View in Article

      Table 1. Data set division

      数据集数量
      训练集1 413
      验证集157
      测试集393
    • Table 2. Definition of small, medium and large goals

      View table
      View in Article

      Table 2. Definition of small, medium and large goals

      类型像素值(area)目标数量
      小目标(small)0<a≤322309
      中目标(medium)322a≤9621 402
      大目标(large)962a2 920
    • Table 3. Training environment

      View table
      View in Article

      Table 3. Training environment

      名称型号、参数
      操作系统Windows10
      CPUi9-10900 CPU@2.80 GHz
      RAM32.0 GB
      显卡GeForce RTX 3070
      DL框架Pytorch1.8
    • Table 4. Training parameters

      View table
      View in Article

      Table 4. Training parameters

      训练方式训练轮次学习率batch_size优化器动量置信度阈值
      冻结训练500.00132Adam0.940.5
      解冻训练2500.000 18Adam0.940.5
    • Table 5. Model results

      View table
      View in Article

      Table 5. Model results

      模型BackboneRecall/%Precision/%AP@0.5/%F1Params/MBFLOPs/GModel Size/kBFPS
      HOG+SVM/40.9769.9948.360.52///2
      FasterRcnnResnet5091.5470.1889.360.7928.275454.292110 77317
      Yolov3Darknet5391.6694.6691.230.9361.52432.759240 69365
      Yolov4Darknet53+SPP93.6592.8992.970.9369.33829.879250 26548
      CenterNetResnet5083.6397.8183.630.9032.66422.125127 93273
      RetinaNetResnet5082.8491.5682.260.8736.35174.042142 25137
      Yolox-sCSPDarknet94.9593.0994.370.948.93813.31735 17767
      IM-YOLO-sIM-CSPDarknet97.1892.6196.770.958.97313.31935 34160
    • Table 6. Ablation experimental results

      View table
      View in Article

      Table 6. Ablation experimental results

      YOLOX-sCA

      Neck

      改进

      Loss

      改进

      AP@0.5/%R/%P/%

      AP@

      0.5:0.95

      APs@

      0.5:0.95

      APm@

      0.5:0.95

      APl@

      0.5:0.95

      FPS/(frame·s-1
      94.3794.9593.090.7750.4380.6880.85767
      95.4896.0093.160.7780.4530.6940.85961
      95.0395.5393.230.7700.4730.6860.84966
      95.2696.0091.490.7620.4430.6820.83867
      95.7196.0093.480.7810.4530.7000.86160
      95.7596.3692.970.7800.4790.6840.86161
      95.2795.6593.140.7870.4290.6910.87666
      96.7797.1892.610.7860.4420.6920.86960
    Tools

    Get Citation

    Copy Citation Text

    Xinwei LIU, Yongjie PIAO, Liangliang ZHENG, Wei XU, Haolin JI. Ship detection for complex scene images of space optical remote sensing[J]. Optics and Precision Engineering, 2023, 31(6): 892

    Download Citation

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

    Category: Information Sciences

    Received: Jun. 8, 2022

    Accepted: --

    Published Online: Apr. 4, 2023

    The Author Email: PIAO Yongjie (pyj0314@163.com)

    DOI:10.37188/OPE.20233106.0892

    Topics