Infrared and Laser Engineering, Volume. 51, Issue 3, 20210106(2022)

Infrared dim and small target detection based on YOLO-IDSTD algorithm

Xinhao Jiang... Wei Cai, Zhiyong Yang, Peiwei Xu and Bo Jiang |Show fewer author(s)
Author Affiliations
  • Armament Launch Theory and Technology Key Discipline Laboratory of PRC, Rocket Force University of Engineering, Xi′an 710025, China
  • show less
    Figures & Tables(16)
    YOLO-IDSTD network structure. (a) Feature extraction part; (b) Feature fusion part; (c) Target detection part
    Structure of Focus
    Structure of PDSCP
    Improved RFB-Small block
    Some images of data set
    Typical infrared dim and small targets in data set
    Comparison of detection results of typical infrared dim and small targets. (a) YOLOv3; (b) YOLOv4-tiny; (c) YOLOv3-tiny; (d) YOLO-IDSTD
    Test results on OSU Thermal Pedistrian Database
    Test results on FLIR Thermal Datasets
    • Table 1. Each layer’s parameters and FLOPs of feature extraction part

      View table
      View in Article

      Table 1. Each layer’s parameters and FLOPs of feature extraction part

      No.NameParameterFLOPs
      1Focus, 1, 16224×10633.0×106
      2Conv, 3/1, 162336×10686.1×106
      3Conv, 3/1, 324672×10643.1×106
      4Conv, 3/1, 6418560×10642.8×106
      5PDSCP, 12838016×10621.9×106
      6PDSCP, 256149760×10621.6×106
      7PDSCP, 512594432×10621.4×106
    • Table 2. Configuration of experimental platform

      View table
      View in Article

      Table 2. Configuration of experimental platform

      NameRelated configurations
      GPUNVIDIA quadro GV100
      CPUsInter Xeon silver 4210/128G
      GPU memory size32G
      Operating systemsWin10
      Computing platformCUDA11.0
      CPU(test)Inter Core i7 10700/16G
    • Table 3. Statistics of extension box

      View table
      View in Article

      Table 3. Statistics of extension box

      Size of extension boxNumber of datasetsNumber of images
      5 pixel×5 pixel1312484
      7 pixel×7 pixel2798
    • Table 4. Setting of experimental parameters

      View table
      View in Article

      Table 4. Setting of experimental parameters

      ParameterInfrared dim and small targets datasetsThermal Pedestrian DatabaseFLIR Thermal Datasets
      Class number113
      Epoch500500500
      Batch size64464
      Image size384×384320×320512×512
      Batch size(test)111
    • Table 5. Precision and efficiency of different detection methods

      View table
      View in Article

      Table 5. Precision and efficiency of different detection methods

      MethodPrecision rateRecall rateAP@0.5ParameterModel size/MBGFLOPsDetection time/ms
      YOLOv3-3840.73710.81820.812361.6×106117.7155.2364.8
      SSD3000.36640.75850.517023.7×10690.635.2370.4
      Mobilenet-SSD0.52410.51110.33006.3×10624.01.1466.8
      Efficientdet b00.59480.05890.09993.9×10615.12.573.8
      Centernet-ResNet500.83230.61560.684332.6×106124.83.830.3
      YOLOv5s-3840.73100.80290.79577.3×10616.617.098.5
      YOLOv4-tiny--3840.67130.78470.81956.2×10612.616.580.1
      YOLOv3-tiny-3840.67800.76520.80508.9×10614.212.978.5
      YOLO-IDSTD0.64050.84090.82423.7×1067.33.050.2
    • Table 6. Ablation experiment of YOLO-IDSTD

      View table
      View in Article

      Table 6. Ablation experiment of YOLO-IDSTD

      Improve the detection speedImprove the detection accuracyRecall rateAP@0.5Model size/MB Detection time/ms
      YOLOv3-tiny baselineWith FocusWith PDSCPWith PANetWith Four-scales predictionWith RFB-Small
      0.78790.777116.5978.5
      0.75760.723516.5935.4
      0.76520.73423.6526.5
      0.76520.76049.1131.7
      0.82580.80379.2236.9
      0.84090.82427.2750.2
    • Table 7. Comparative experiments on two sets of infrared small target datasets

      View table
      View in Article

      Table 7. Comparative experiments on two sets of infrared small target datasets

      MethodOSU Thermal Pedestrian DatabaseFLIR Thermal Datasets
      Recall rateAP@0.5Detection time/ ms Recall ratemAP@0.5AP@0.5 (person) AP@0.5 (bicycle) AP@0.5 (car) Detection time/ ms
      Efficientdet b00.87230.865190.50.33740.49430.4440.4350.604160.8
      YOLOv5s0.99090.986069.60.77060.74410.7990.5630.870122.6
      YOLOv3-tiny10.987553.20.69060.63340.6410.4490.81098.4
      YOLO-IDSTD10.989942.90.71660.66760.7240.4480.83160.7
    Tools

    Get Citation

    Copy Citation Text

    Xinhao Jiang, Wei Cai, Zhiyong Yang, Peiwei Xu, Bo Jiang. Infrared dim and small target detection based on YOLO-IDSTD algorithm[J]. Infrared and Laser Engineering, 2022, 51(3): 20210106

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category: Image processing

    Received: Feb. 20, 2021

    Accepted: --

    Published Online: Apr. 8, 2022

    The Author Email:

    DOI:10.3788/IRLA20210106

    Topics