• Journal of Infrared and Millimeter Waves
  • Vol. 43, Issue 4, 582 (2024)
Rui ZHANG1, Min LIU1, and Zheng LI2,*
Author Affiliations
  • 1School of Opto-Electronic and Comunication Engineering,Xiamen University of Technology,Xiamen 361024,China
  • 2Shanghai Institute of Technical Physics,Chinese Academy of Sciences,Shanghai 200083,China
  • show less
    DOI: 10.11972/j.issn.1001-9014.2024.04.019 Cite this Article
    Rui ZHANG, Min LIU, Zheng LI. Research on fast detection method of infrared small targets under resource-constrained conditions[J]. Journal of Infrared and Millimeter Waves, 2024, 43(4): 582 Copy Citation Text show less
    Infrared small UAVs
    Fig. 1. Infrared small UAVs
    Improved YOLOv5s network architecture
    Fig. 2. Improved YOLOv5s network architecture
    Small target detection head
    Fig. 3. Small target detection head
    The sensitivity analysis of IoU on infrared small UAV
    Fig. 4. The sensitivity analysis of IoU on infrared small UAV
    Dataset analysis:(a) the label category distribution; (b) the bounding box size distribution; (c) the label center position distribution; (d) the label size distribution
    Fig. 5. Dataset analysis:(a) the label category distribution; (b) the bounding box size distribution; (c) the label center position distribution; (d) the label size distribution
    Performance comparison of the AP:(a) AP@0.5; (b) AP@0.5:0.95
    Fig. 6. Performance comparison of the AP:(a) AP@0.5; (b) AP@0.5:0.95
    Some examples of the detection result on the improved model
    Fig. 7. Some examples of the detection result on the improved model
    Framework of the deployment
    Fig. 8. Framework of the deployment
    Weighting Coefficients

    AP@0.5

    (%)

    AP@0.5:0.95

    (%)

    FAR

    (%)

    MR

    (%)

    LP2+LP3+LP4+0.LP581.244.96.428.4
    LP2+LP3+LP4+0.LP586.847.64.719.4
    LP2+LP3+LP4+0.LP584.246.07.822.8
    LP2+LP3+LP4+0.LP588.448.64.017.4
    Table 1. Comparison of the different weighting coefficient results
    ModelsAP@0.5(%)

    AP@0.5:0.95

    (%)

    FAR

    (%)

    MR

    (%)

    YOLOv5s84.746.13.123.6
    YOLOv5s+0.5NWD87.447.94.017.4
    YOLOv5s+NWD89.948.14.414.6
    YOLOv5s+P288.448.64.017.4
    YOLOv5s+NWD+P291.950.04.212.9
    Table 2. Comparison of ablation experiments of improved methods
    Models

    AP@0.5

    (%)

    AP@0.5:0.95

    (%)

    FAR

    (%)

    MR

    (%)

    Parameter

    (M)

    GFLOPs

    Speed

    (FPS)

    Weights

    (MB)

    SSD-ResNet5060.422.829.746.213.115.0200105.1
    Faster-RCNN-ResNet5078.330.921.240.041.1134.550330.3
    RetinaNet-ResNet5082.133.818.533.232.0127.543257.3
    YOLOv383.345.28.722.49.323.152618.9
    YOLOv5s84.746.13.123.67.015.862514.4
    YOLOv5m86.648.83.120.420.847.930342.2
    YOLOv5l87.749.13.817.646.1107.619692.8
    YOLOv8s89.548.96.317.311.128.443522.5
    YOLOv5s+NWD+P291.950.04.212.97.726.840016.3
    Table 3. Comparison of improved YOLOv5s with other methods
    BM1684XFPSAP@0.5 (%)AP@0.5:0.95(%)
    FP321291.950.0
    FP169587.849.8
    INT8163--
    Table 4. Result of the deployment
    Rui ZHANG, Min LIU, Zheng LI. Research on fast detection method of infrared small targets under resource-constrained conditions[J]. Journal of Infrared and Millimeter Waves, 2024, 43(4): 582
    Download Citation