• Acta Photonica Sinica
  • Vol. 51, Issue 9, 0910001 (2022)
Yinhui ZHANG, Pengcheng ZHANG, Zifen HE*, and Sen WANG
Author Affiliations
  • Mechanical and Electrical Engineering,Kunming University of Science and Technology,Kunming 650500,China
  • show less
    DOI: 10.3788/gzxb20225109.0910001 Cite this Article
    Yinhui ZHANG, Pengcheng ZHANG, Zifen HE, Sen WANG. Lightweight Real-time Detection Model of Infrared Pedestrian Embedded in Fine-scale[J]. Acta Photonica Sinica, 2022, 51(9): 0910001 Copy Citation Text show less
    Visualization of the feature map
    Fig. 1. Visualization of the feature map
    Structure of TIPRD network
    Fig. 2. Structure of TIPRD network
    The distribution of γ before and after sparse training
    Fig. 3. The distribution of γ before and after sparse training
    Comparison of channel numbers before and after pruning
    Fig. 4. Comparison of channel numbers before and after pruning
    Model compression framework
    Fig. 5. Model compression framework
    Example of data
    Fig. 6. Example of data
    Test results
    Fig. 7. Test results
    Detector sizeTarget sizeAnchor size
    16×16Large target

    (38,132)

    (62,205)

    (116,321)

    32×32Medium target

    (13,49)

    (19,60)

    (25,89)

    64×64Small target

    (5,17)

    (8,26)

    (9,39)

    Table 1. The allocation strategy of anchor
    EquipmentModelQuantity
    CPUIntel(R)Core(TM)i5-10400F1
    GPUNVIDIA GeForce RTX 2080Ti1
    RAM8G DDR4 26662
    Hard disk512G SSD1
    Table 2. Hardware configuration
    ParameterValue
    Learn rate0.001
    Epoch600
    Momentum0.9
    Hue0.1
    PolicySteps
    Steps3 200,3 600
    Table 3. Parameter configuration
    ModelFine⁃scale detection layer4×CSPK⁃means++

    Size/

    MB

    Speed/

    (frame·s-1

    mAP/

    %

    Yolov4⁃tiny23.590.780.6
    Our24.984.086.5
    Our24.979.987.5
    Yolo⁃pedestrian24.979.389.2
    Table 4. Comparison of the effect of precision improvement strategies
    Pruning rateFine tuningKnowledge distillationNumber of channelsSize/MBmAP/%
    0.002 91224.989.2
    0.757284.689.1
    0.757284.689.5
    0.805834.088.6
    0.805834.089.2
    0.853743.587.7
    0.853743.588.0
    0.901843.083.9
    0.901843.084.1
    Table 5. Comparison of results of different fine-tuning strategies
    ModelSize/MBSpeed/(frame·s-1mAP/%
    Yolov3246.349.491.2
    Yolov3-tiny34.794.279.7
    Yolov425643.391.6
    Yolov4-tiny23.590.780.6
    Yolo-pedestrian24.979.389.2
    TIPRD4.088.789.2
    Table 6. Comparison of different algorithms
    ModelSpeed/(frame·s-1
    Yolov4/
    Yolov30.95
    Yolov4-tiny5.2
    Yolov3-tiny2.6
    TIPRD6.9
    Table 7. Comparison of detection speed of different models
    Yinhui ZHANG, Pengcheng ZHANG, Zifen HE, Sen WANG. Lightweight Real-time Detection Model of Infrared Pedestrian Embedded in Fine-scale[J]. Acta Photonica Sinica, 2022, 51(9): 0910001
    Download Citation