• Optics and Precision Engineering
  • Vol. 30, Issue 13, 1591 (2022)
Zifen HE, Guangchen CHEN, Sen WANG, Yinhui ZHANG*, and Linwei GUO
Author Affiliations
  • Faculty of Mechanical and Electrical Engineering, Kunming University of Science and Technology, Kunming650500, China
  • show less
    DOI: 10.37188/OPE.20223013.1591 Cite this Article
    Zifen HE, Guangchen CHEN, Sen WANG, Yinhui ZHANG, Linwei GUO. Detection of foreign object debris on night airport runway fusion with self-attentional feature embedding[J]. Optics and Precision Engineering, 2022, 30(13): 1591 Copy Citation Text show less
    CTPNet Network Structure
    Fig. 1. CTPNet Network Structure
    Comparison of bottleneck structures
    Fig. 2. Comparison of bottleneck structures
    Multi-head self attention structure
    Fig. 3. Multi-head self attention structure
    Fitting result of real box and prediction box
    Fig. 4. Fitting result of real box and prediction box
    CIoU loss of signal
    Fig. 5. CIoU loss of signal
    NFOD dataset images and annotations
    Fig. 6. NFOD dataset images and annotations
    Target instance scale distribution
    Fig. 7. Target instance scale distribution
    Visualization results of mean average precision
    Fig. 8. Visualization results of mean average precision
    Test result visualization
    Fig. 9. Test result visualization
    Visualization of characteristic image
    Fig. 10. Visualization of characteristic image
     
    传感器规格高级COMS感光芯片 1/2.7 inch
    像元尺寸3 μm×3 μm
    最低工作照度0.051 lx
    速度30 frame/s
    输出分辨率1 280×720
    Table 1. Parameters of LRCP20680_1080P camera
    检测层聚类前聚类后
    20×20(10, 13), (16,30), (33, 23)(6, 8), (10,15), (12, 24)
    40×40(30, 61), (62, 45), (59, 119)(16, 18), (22, 27), (33, 16)
    80×80(116, 90), (156, 198), (373, 326)(37, 77), (42, 35), (66, 68)
    Table 2. Initial candidate box size of detect layers
    模 型GIoUK-meansCIoUTransformer BottleNeck

    Weight

    /MB

    Speed/

    (frame·s-1

    mAP

    /%

    YOLOv5+GIoU14.441.882.9
    YOLOv5+K-means+ GIoU14.443.283.6
    YOLOv5+K-means+CIoU14.442.584.3
    YOLOv5+K-means+CIoU+TransformerBotteNeck14.438.088.1
    Table 3. Result of ablation experiments
    Model

    Speed/

    (frame·s-1

    Weight

    /MB

    mAP

    (%)

    Plier

    (%)

    Screwdriver

    (%)

    Strapping_tape

    (%)

    Nail

    (%)

    Sheetmetal

    (%)

    Spanner

    (%)

    Branch

    (%)

    Nut

    (%)

    Block_rubber

    (%)

    CSPTNet-1H41.514.487.291.880.783.881.690.892.075.689.387.2
    CSPTNet-2H39.414.487.281.791.085.778.587.290.074.498.098.1
    CSPTNet-4H38.014.488.186.993.377.774.095.694.483.596.191.5
    CSPTNet-8H28.514.487.182.586.092.383.694.090.166.289.987.1
    CSPTNet-16H20.614.484.179.686.181.280.491.194.576.682.584.7
    Table 4. Comparison of effect of subspace number of self-attentional branches
    Model

    Speed/

    (frame·s-1

    Weight

    /MB

    mAP

    (%)

    Plier

    (%)

    Screwdriver

    (%)

    Strapping_tape

    (%)

    Nail

    (%)

    Sheetmetal

    (%)

    Spanner

    (%)

    Branch

    (%)

    Nut

    (%)

    Block_rubber

    (%)

    SE40.61576.577.162.870.577.590.885.472.257.594.5
    CoordAtt39.014.577.670.374.581.679.489.778.656.676.290.6
    CBAM42.014.579.967.182.480.477.492.586.365.875.591.3
    ChannleAtt45.014.580.766.775.781.484.188.590.071.778.989.0
    ECA42.714.883.769.185.783.781.489.887.872.684.598.6
    SAM41.814.585.485.986.182.582.985.895.176.482.491.5
    MHSA38.014.488.186.993.377.774.095.694.483.596.191.5
    Table 5. Comparative experiment results of attention mechanism
    ModelSpeed/(frame·s-1

    Weight

    /MB

    mAP

    (%)

    Plier

    (%)

    Screwdriver

    (%)

    Strapping_tape

    (%)

    Nail

    (%)

    Sheetmetal

    (%)

    Spanner

    (%)

    Branch

    (%)

    Nut

    (%)

    Block_rubber

    (%)

    YOLOv5-ST41.214.782.466.482.377.076.187.490.782.986.692.0
    YOLOv5-Ghost47.513.283.575.385.882.580.590.491.183.772.789.8
    YOLOv5-CSP43.914.685.090.386.787.078.188.989.170.385.488.9
    CSPTNet38.014.488.186.993.377.774.095.694.483.596.191.5
    Table 6. Effect comparison of bottleneck modules
    ModelSpeed/(frame·s-1

    Weight

    /MB

    mAP

    (%)

    Plier

    (%)

    Screwdriver

    (%)

    Strapping_tape

    (%)

    Nail

    (%)

    Sheetmetal

    (%)

    Spanner

    (%)

    Branch

    (%)

    Nut

    (%)

    Block_rubber

    (%)

    YOLOv3-tiny49.717.430.340.59.026.422.259.342.4013.559.1
    VarifocalNet14.9261.452.870.756.769.52.842.878.275.31.477.4
    Faster R-CNN19.9330.665.688.972.087.320.753.785.480.621.780.5
    Sparse R-CNN17.2130073.085.365.793.847.172.879.179.359.274.5
    TOOD16.6255.875.184.081.590.049.862.390.980.960.081.8
    YOLOx14.871.978.6992.681.597.556.187.898.083.323.887.6
    YOLOv339.519.482.959.981.388.671.194.587.775.396.591.5
    YOLOv541.814.482.977.872.482.576.688.589.882.376.299.5
    Ours38.014.488.186.993.377.77495.694.483.596.191.5
    Table 7. Comparison of model effects
    Zifen HE, Guangchen CHEN, Sen WANG, Yinhui ZHANG, Linwei GUO. Detection of foreign object debris on night airport runway fusion with self-attentional feature embedding[J]. Optics and Precision Engineering, 2022, 30(13): 1591
    Download Citation