• Optics and Precision Engineering
  • Vol. 32, Issue 12, 1929 (2024)
Meng HE, Jiangpeng WU*, Chao LIANG, Pengyu HU..., Yuan REN, Xuan HE and Qianghui LIU|Show fewer author(s)
Author Affiliations
  • Xi’an Modern Control Technology Research Institute, Xi’an710065, China
  • show less
    DOI: 10.37188/OPE.20243212.1929 Cite this Article
    Meng HE, Jiangpeng WU, Chao LIANG, Pengyu HU, Yuan REN, Xuan HE, Qianghui LIU. Few-shot warhead fragment group object detection based on feature reassembly and attention[J]. Optics and Precision Engineering, 2024, 32(12): 1929 Copy Citation Text show less
    Network structure of YOLOv5s-FD model
    Fig. 1. Network structure of YOLOv5s-FD model
    CARAFE module structure
    Fig. 2. CARAFE module structure
    CA module structure
    Fig. 3. CA module structure
    Framework for training object detection model using MAML method
    Fig. 4. Framework for training object detection model using MAML method
    Backbone network freeze layer
    Fig. 5. Backbone network freeze layer
    Self-made fragment dataset presentation
    Fig. 6. Self-made fragment dataset presentation
    Comparison of precision and recall metrics for different models
    Fig. 7. Comparison of precision and recall metrics for different models
    Comparison of average precision metrics for different models
    Fig. 8. Comparison of average precision metrics for different models
    Comparison of YOLOv5s and this paper model detection effect scene 1
    Fig. 9. Comparison of YOLOv5s and this paper model detection effect scene 1
    Comparison of YOLOv5s and this paper model detection effect scene 2
    Fig. 10. Comparison of YOLOv5s and this paper model detection effect scene 2
    ProjectEnvironment
    CPUIntel(R) Xeon(R) Silver 4215R CPU @ 3.20 GHz 3.19 GHz
    Memory128 GB
    Operating SystemUbuntu 20.04
    GPU4 NVIDIA RTX3090
    Python versionPython3.9
    Pytorch versionPyTorch1.10
    Object Detection FrameworkYOLOv5s-FD
    GPU Acceleration LibraryCUDA cuDNN
    Table 1. Experimental environment
    Training MethodPRmAP_0.5mAP_0.5∶0.95
    Traditional0.8620.7930.8450.546
    MAML0.9050.8540.8820.611
    Table 2. Comparative experiment on the effectiveness of MAML meta learning training methods
    YOLOv5sSODLCARAFECAPRmAP_0.5mAP_0.5∶0.95

    Size

    /MB

    Params

    /M

    GFLOPs

    /G

    FPS(frames/s)
    10.8340.7750.8070.54214.47.0215.9127
    20.8690.8130.8370.57614.87.1718.3114
    30.8780.8260.8510.58815.17.3219.4104
    40.9050.8540.8820.61115.67.9420.195
    Table 3. Ablation experimental results
    ModelPRmAP_0.5mAP_0.5:0.95

    Size

    /MB

    Params

    /M

    GFLOPs

    /G

    FPS(frames/s)
    Faster R-CNN0.6990.5680.598-108.9136.60402.329
    SSD0.6300.4950.518-100.324.40274.181
    YOLOv5s0.8340.7750.8070.54214.47.0215.9127
    YOLOv8s0.8260.7890.8170.58122.511.1328.6112
    YOLOv5s-FD(ours)0.9050.8540.8820.61115.67.9420.195
    Table 4. Performance comparison of different algorithm models on Fragment data set
    Meng HE, Jiangpeng WU, Chao LIANG, Pengyu HU, Yuan REN, Xuan HE, Qianghui LIU. Few-shot warhead fragment group object detection based on feature reassembly and attention[J]. Optics and Precision Engineering, 2024, 32(12): 1929
    Download Citation