Laser & Optoelectronics Progress, Volume. 58, Issue 6, 610012(2021)

Small Target Detection Based on Cross-Scale Fusion Convolution Neural Network

Liu Feng1,2, Guo Meng1,2, and Wang Xiangjun1,2
Author Affiliations
  • 1State Key Laboratory Precision Measuring Technology and Instruments, Tianjin University, Tianjin 300072, China
  • 2Micro Optics Electronic Machine System Education Ministry Key Laboratory, Tianjin University, Tianjin 300072, China
  • show less
    Figures & Tables(18)
    Structure of the Darknet-53
    DOTA data set. (a) Original image; (b) cropped image
    Enhancement of the DOTA data set. (a) Original image; (b) enhanced image
    Improved network prediction structure1
    Improved network prediction structure2
    Optimized network of the receptive field
    Detection effect of different networks. (a) YOLOv3; (b) structure1; (c) structure2
    Loss curve during training
    Recognition effect of different networks. (a) Structure2; (b) optimize the network of the receptive field
    Detection results of different networks under the COCO data set. (a) YOLOv3 network; (b) optimize the network of the receptive field
    • Table 1. Performance of different backbone networks

      View table

      Table 1. Performance of different backbone networks

      BackboneTop-1/%Top-5/%FPS /frame
      Darknet-1974.191.8171
      ResNet-10177.193.753
      ResNet-15277.693.837
      Darknet-5377.293.878
    • Table 2. Recall rates of different networks unit: %

      View table

      Table 2. Recall rates of different networks unit: %

      NetworkPlaneLarge-vehicleSmall-vehicleAverage
      YOLOv398.085.882.488.6
      Improved structure198.390.085.691.3
      Improved structure297.888.996.994.5
    • Table 3. Precision rates of different networks unit: %

      View table

      Table 3. Precision rates of different networks unit: %

      NetworkPlaneLarge-vehicleSmall-vehicleAverage
      YOLOv397.480.081.486.3
      Improved structure197.083.085.088.3
      Improved structure296.581.187.188.2
    • Table 4. Multi-category recall rates of different networks unit: %

      View table

      Table 4. Multi-category recall rates of different networks unit: %

      NetworkPlaneLarge-vehicleSmall-vehicleAverage
      YOLOv364.327.511.634.5
      R-FCN72.931.914.239.7
      Improved structure289.657.261.069.3
      Receptive field optimization93.176.872.780.9
    • Table 5. Multi-class precision rates of different networks unit: %

      View table

      Table 5. Multi-class precision rates of different networks unit: %

      NetworkPlaneLarge-vehicleSmall-vehicleAverage
      YOLOv362.016.73.527.4
      R-FCN68.629.113.534.9
      Improved structure287.945.844.859.5
      Receptive field optimization90.337.749.459.0
    • Table 6. Basic parameters of different networks

      View table

      Table 6. Basic parameters of different networks

      NetworkVolume /MbTime consuming /s
      YOLOv3246.30.063
      R-FCN102.50.180
      Improved structure2242.70.085
      Receptive field optimization239.40.083
    • Table 7. Recall rates of different networks under the COCO data set unit: %

      View table

      Table 7. Recall rates of different networks under the COCO data set unit: %

      NetworkSmallMediumLargeAverage
      YOLOv324.048.261.144.4
      Receptive field optimization36.258.265.553.3
    • Table 8. Precision rates of different networks under the COCO data set unit: %

      View table

      Table 8. Precision rates of different networks under the COCO data set unit: %

      NetworkSmallMediumLargeAverage
      YOLOv314.234.146.431.6
      Receptive field optimization25.241.548.538.4
    Tools

    Get Citation

    Copy Citation Text

    Liu Feng, Guo Meng, Wang Xiangjun. Small Target Detection Based on Cross-Scale Fusion Convolution Neural Network[J]. Laser & Optoelectronics Progress, 2021, 58(6): 610012

    Download Citation

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

    Category: Image Processing

    Received: Jul. 17, 2020

    Accepted: --

    Published Online: Mar. 2, 2021

    The Author Email:

    DOI:10.3788/LOP202158.0610012

    Topics