Optics and Precision Engineering, Volume. 31, Issue 16, 2465(2023)

Remote sensing multi-scale object detection based on multivariate feature extraction and characterization optimization

Yuebo MENG... Fei WANG, Guanghui LIU* and Shengjun XU |Show fewer author(s)
Author Affiliations
  • College of Information and Control Engineering, Xi 'an University of Architecture and Technology, Xi'an710055, China
  • show less
    Figures & Tables(20)
    Network structure of multivariate feature extraction and characterization optimization
    MFE module structure
    CM module structure
    Example of remote sensing large aspect ratio target
    Label optimization comparison
    CPA module structure
    Some examples images from DIOR dataset
    Experiential results of our proposed method(shown on the down) and Baseline(shown on the up) on DIOR dataset
    Example images from HRRSD dataset
    Experiential results of our proposed method(shown on the down) and Baseline(shown on the up) on HRRSD dataset
    Example images from RSOD dataset
    Experiential results of our proposed method(shown on the down) and Baseline(shown on the up) on RSOD dataset
    • Table 1. Performance comparison of different algorithms on DIOR dataset

      View table
      View in Article

      Table 1. Performance comparison of different algorithms on DIOR dataset

      Faster RCNN1

      Mask

      R-CNN2

      YOLOx3Retina-Net4PANet31CSFF7CBD-E32Corner-Net14Baseline

      本文方法

      MFC

      D760.060.453.162.461.462.763.564.3*51.454.5
      D875.676.355.678.672.182.6*76.281.662.378.1
      D962.362.557.262.866.773.265.376.3*60.067.8
      D1076.076.077.578.672.078.279.379.575.380.5*
      D1176.875.973.376.673.481.6*79.579.574.076.6
      D1246.446.548.249.945.350.747.526.151.354.3*
      D1357.257.453.759.656.959.559.360.655.861.5*
      D1471.871.883.971.171.773.369.137.684.787.4*
      D1568.368.362.268.470.463.469.770.7*56.265.3
      D1653.853.761.445.862.058.564.345.268.972.6*
      D1781.181.086.381.380.985.984.584.086.388.2*
      D1859.562.361.455.257.061.959.457.153.462.4*
      D1943.143.047.344.447.242.944.743.045.147.9*
      D2081.281.084.585.584.586.983.175.975.787.2*
      mAP/%↑65.165.265.966.166.168.067.864.966.270.9*
      FPS↑11.212.933.216.414.5--26.730.528.1
      FLOPs↓38.7 G38.4 G11.4 G25.2 G26.9 G--12.8 G12.5 G12.7 G
    • Table 1. Performance comparison of different algorithms on DIOR dataset

      View table
      View in Article

      Table 1. Performance comparison of different algorithms on DIOR dataset

      Faster RCNN1

      Mask

      R-CNN2

      YOLOx3Retina-Net4PANet31CSFF7CBD-E32Corner-Net14Baseline

      本文方法

      MFC

      D154.053.972.553.360.257.254.258.878.187.4*
      D274.576.676.377.072.079.677.084.2*71.283.6
      D363.363.274.169.370.670.171.572.072.674.3*
      D480.780.977.485.080.587.487.180.887.187.8*
      D544.840.240.844.143.646.144.646.438.847.2*
      D672.572.570.573.272.376.675.475.375.377.3*
    • Table 2. Performance comparison of different algorithms on HRRSD dataset

      View table
      View in Article

      Table 2. Performance comparison of different algorithms on HRRSD dataset

      Faster RCNN1

      GACL

      FasterRCNN29

      YOLOx3RetinaNet4PANet31CornerNet14Baseline

      本文方法

      MFC

      H190.590.891.392.693.788.596.698.4*
      H286.188.585.686.787.289.290.792.2*
      H387.389.286.789.190.593.394.195.0*
      H486.787.286.387.589.389.588.491.1*
      H580.780.885.384.987.790.691.893.8*
      H648.249.771.751.358.272.671.475.7*
      H789.990.787.489.692.394.094.296.9*
      H890.189.790.289.989.789.991.594.8*
      H986.285.685.186.288.488.588.792.3*
      H1086.586.985.487.788.989.290.393.8*
      H1189.488.287.587.489.490.891.994.9*
      H1274.775.074.175.379.379.480.085.8*
      H1366.265.365.368.3*66.566.863.267.6
      mAP/%↑81.782.183.282.884.786.387.190.2*
      FPS↑14.5-48.923.220.942.245.543.7
      FLOPs↓38.7 G-11.4 G25.2 G26.9 G12.8 G12.5 G12.7 G
    • Table 3. Performance comparison of different algorithms on RSOD dataset

      View table
      View in Article

      Table 3. Performance comparison of different algorithms on RSOD dataset

      Faster RCNN1

      Mask

      R-CNN2

      YOLOx3RetinaNet4CBD-E32CornerNet14Baseline本文方法MFC
      R179.681.287.582.295.8*92.191.494.4
      R294.793.495.399.799.796.896.399.7*
      R385.085.983.693.888.890.090.194.1*
      R493.893.591.297.194.293.594.999.5*
      mAP/%↑88.388.589.493.294.293.193.296.9*
      FPS↑15.917.561.327.6-56.361.158.8
      FLOPs↓38.738.4G11.4G25.2G-12.8G12.5G12.7G
    • Table 4. Ablation experiments on RSOD datasets(MFE)

      View table
      View in Article

      Table 4. Ablation experiments on RSOD datasets(MFE)

      MethodmAP/%↑FPS↑FLOPs↓
      Baseline93.261.512.48 G
      +跨组连接94.959.912.64 G
      +跨组连接+SE95.259.712.69 G
      +跨组连接+SE+CM95.759.412.71 G
    • Table 5. Ablation experiments on RSOD datasets(COS)

      View table
      View in Article

      Table 5. Ablation experiments on RSOD datasets(COS)

      MethodmAP/%↑FPS↑FLOPs↓
      Baseline93.261.512.48 G
      +标签优化93.561.412.49 G
      +标签优化+CPA94.160.112.52 G
    • Table 6. Contrast experiments on RSOD datasets(SE、CA)

      View table
      View in Article

      Table 6. Contrast experiments on RSOD datasets(SE、CA)

      MethodmAP/%↑FPS↑FLOPs↓
      Baseline93.261.512.48 G
      +MFE(with SE)95.759.412.71 G
      +MFE(with CBAM)95.859.412.72 G
      +COS(with CA2694.160.112.52 G
      +COS(with CA2793.960.212.50 G
    • Table 7. Ablation experiments on DIOR、HRRSD and RSOD datasets

      View table
      View in Article

      Table 7. Ablation experiments on DIOR、HRRSD and RSOD datasets

      消融实验DIORHRRSDRSODFLOPs
      MFECOSmAP/%FPSMFECOSmAP/%FPSMFECOSmAP/%FPS
      Baseline--66.230.5--87.145.5--93.261.512.5 G
      -69.329.2-89.443.1-95.759.412.7 G
      70.928.190.243.796.958.812.7 G
    Tools

    Get Citation

    Copy Citation Text

    Yuebo MENG, Fei WANG, Guanghui LIU, Shengjun XU. Remote sensing multi-scale object detection based on multivariate feature extraction and characterization optimization[J]. Optics and Precision Engineering, 2023, 31(16): 2465

    Download Citation

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

    Category: Information Sciences

    Received: Nov. 10, 2022

    Accepted: --

    Published Online: Sep. 5, 2023

    The Author Email: LIU Guanghui (guanghuil@163.com)

    DOI:10.37188/OPE.20233116.2465

    Topics