• Laser & Optoelectronics Progress
  • Vol. 60, Issue 2, 0228004 (2023)
Lei Lang1, Kuan Liu2, and Dong Wang1,*
Author Affiliations
  • 1School of Computer and Information Technology, Beijing Jiaotong University, Beijing 100044, China
  • 2Computer and Communication Engineering, Zhengzhou University of Light Industry, Zhengzhou 450001, Henan , China
  • show less
    DOI: 10.3788/LOP212699 Cite this Article Set citation alerts
    Lei Lang, Kuan Liu, Dong Wang. Lightweight Remote Sensing Object Detector based on YOLOX-Tiny[J]. Laser & Optoelectronics Progress, 2023, 60(2): 0228004 Copy Citation Text show less
    YOLOX-Tiny network structure
    Fig. 1. YOLOX-Tiny network structure
    Mapping of anchor in original image at different scales
    Fig. 2. Mapping of anchor in original image at different scales
    Improved YOLOX-Tiny
    Fig. 3. Improved YOLOX-Tiny
    Coordinate attention module
    Fig. 4. Coordinate attention module
    Deformable convolution schematic diagram
    Fig. 5. Deformable convolution schematic diagram
    Images and objects in DIOR dataset
    Fig. 6. Images and objects in DIOR dataset
    Size distribution of objects in DIOR dataset by category
    Fig. 7. Size distribution of objects in DIOR dataset by category
    AP curve during training
    Fig. 8. AP curve during training
    Loss curve during training
    Fig. 9. Loss curve during training
    Thermal map visualization results
    Fig. 10. Thermal map visualization results
    Comparison of detection results between YOLOX-Tiny and optimized model. (a) YOLOX-Tiny;(b) proposed algorithm
    Fig. 11. Comparison of detection results between YOLOX-Tiny and optimized model. (a) YOLOX-Tiny;(b) proposed algorithm
    NameValue
    OptimizerSGD
    Momentum0.9
    Weight decay5×10-4
    NesterovTrue
    Learning rate scheduler

    Type is CosineAnealing,

    Learning rate is 0.0025,

    Min_lr_ratio is 0.05

    Batch16
    Epoch100
    MosaicImg_scale is (640,640)
    Random affineScaling_ratio_range is (0.5,1.5)
    Photometric distortion

    Brightness_delta is 32,

    Contrast_range is (0.5, 1.5),

    Saturation_range is (0.5, 1.5),

    Hue_delta is 18

    Table 1. Training parameters
    MethodParameters /106AP /%AP50 /%
    YOLOX-Tiny5.0446.071.66
    + multi-scale prediction method5.4047.4(+1.4)75.10(+3.44)
    + coordinate attention module5.4147.7(+0.3)75.30(+0.20)
    + deformable convolution5.649.8(+2.1)75.60(+0.30)
    +loss function(proposed optimized model)5.650.1(+0.3)76.08(+0.48)
    Table 2. Comparison of progressively improved algorithms
    MethodAPAPSAPMAPLAP50AP50SAP50MAP50L
    YOLOX-Tiny46.09.535.766.371.6624.158.590.4
    Proposed optimized model50.112.838.670.276.0831.563.191.8
    Table 3. Improved algorithm detection results under different scales
    MethodC1C2C3C4C5C6C7C8C9C10C11C12C13C14C15C16C17C18C19C20
    YOLOX-Tiny6884.679.986.642.678.371.386.566.180.878.861.559.879.769.861.587.567.440.182.3
    Proposed optimized model74.489.183.888.347.278.677.588.876.182.481.264.562.287.774.370.988.870.349.186.4
    Table 4. AP50 results under different categories
    MethodYearBackboneParameters /106FLOPs /109AP50 /%FPSDevice
    CF2PN342021VGG1691.6>3167.2519.7RTX 2080
    ASSD352021VGG16>40>3171.821RTX TITAN
    LO-Det362020MobileNetv26.936.42465.8560.03RTX 3090
    Proposed optimized model2021Modified CSPNet5.67.69576.0846.0RTX 2080Ti
    Table 5. Comparison of results in DIOR dataset