Infrared and Laser Engineering, Volume. 51, Issue 12, 20220253(2022)

Lightweight infrared dim vehicle target detection algorithm based on deep learning

Renhao Cai1, Ning Cheng1, Zhiyong Peng1, Shize Dong1, Jianmin An2, and Gang Jin3
Author Affiliations
  • 1Tianjin Jinhang Institute of Technical Physics, Tianjin 300308, China
  • 2The Third Military Representative Office in Tianjin, Tianjin 300308, China
  • 3Tianjin University, Tianjin 300072, China
  • show less
    Figures & Tables(22)
    Development trend of deep learning object detection algorithms in recent years
    Effect drawing of data expansion
    Structure diagram of JH-YOLOv5-RDAB
    Structure of YOLOv5 s
    Structure diagram of SPP module after merging EPA module
    Structure diagram of CSP module after merging EPA module
    Structure diagram of ECA module
    Structure diagram of EPA module
    Visible light image (left) and infrared image (right) of the same scene
    Structure diagram of Focus module after merging RDAB module
    Structure diagram of SPP module after merging RDAB module
    Structure diagram of RDAB
    Structure diagram of RDB
    Image comparison of training results
    Image comparison of some detection results (left: actual annotation map, right: algorithm detection map)
    • Table 1. Analog image processing method for single image

      View table
      View in Article

      Table 1. Analog image processing method for single image

      Flight live imageAnalog image processing
      Aircraft falling at high speedImage magnification under fixed field of view
      Aircraft level flightImage translation
      Aircraft rotationImage rotation at various angles
      Aircraft shakingImage translation
      Infrared imagers are affected by temperature and weatherImage contrast, brightness changes
      Aerodynamic effect of aircraft flying at high speedImage random blur, edge blur
      Aircraft is disturbedImage random occlusion
    • Table 2. Distribution of infrared small and weak vehicle image dataset

      View table
      View in Article

      Table 2. Distribution of infrared small and weak vehicle image dataset

      Target type[car]
      Target sceneDesert, city, field, highway, village
      Data Augmentation MethodBrightness change, contrast change, rotation, translation, scaling, flipping, clipping, splicing
      Original data Set5023 (Before treatment) 4986 (After treatment)
      Data augmentation1000
      Training set49228
      Validation set2590
    • Table 3. Specific parameters of the experimental en-vironment

      View table
      View in Article

      Table 3. Specific parameters of the experimental en-vironment

      Experimental systemUbuntu18.04
      CPUInter Xeon Gold 6133
      GPUNVIDIA TITAN RTX ×4
      Memory512 GB
      Development environmentPython3.7
      Deep learning frameworkPytorch
      CUDA10.2
      cuDNN7.5.3
    • Table 4. Comparison of experimental results between JH-YOLOv5-RDAB and typical algorithm networks

      View table
      View in Article

      Table 4. Comparison of experimental results between JH-YOLOv5-RDAB and typical algorithm networks

      AlgorithmLayersParametersSize (Semi precision quantization)mAP50GFLOPsSingle test time/ms
      YOLOv326161497430123.4M81.6154.95.3
      YOLOv4-tiny99587411623.6M88.216.11.8
      YOLOv448863937686256.3M94.6141.48.7
      YOLOv5 s283706354214.4M93.416.32.34
      YOLOv5 m3912105640642.5M94.550.33.45
      YOLOv5 l4994663135093.7M95.2114.14.93
      YOLOv5 x60787244374175.1M95.9217.18.43
      YOLO-Fastest2773567004.8M32.10.961.6
      JH-YOLOv5-RDAB56531173136.6M95.18.42.52
    • Table 5. Comparison of experimental effects of lightweight tailoring based on YOLOv5 s

      View table
      View in Article

      Table 5. Comparison of experimental effects of lightweight tailoring based on YOLOv5 s

      AlgorithmDepth_ multiple width_ multiple Performance effect/timesNetwork size/times
      YOLOv5 x1.331.252.312.2
      YOLOv5 l111.86.5
      YOLOv5 m0.670.751.22.9
      YOLOv5 s0.330.5011
      Net 10.250.330.850.49
      Net 20.20.250.630.35
      Net 30.10.20.330.12
    • Table 6. Comparison of network experimental effects based on attention mechanism fusion

      View table
      View in Article

      Table 6. Comparison of network experimental effects based on attention mechanism fusion

      Fused operatorIntegrated ECAIntegrated EPA
      Performance effect/timesNetwork size/timesPerformance effect/timesNetwork size/times
      Net 21111
      CBL1.11.61.32.4
      SPP1.21.121.431.18
      CSP1.31.11.331.11
      Focus1.021.011.111.01
    • Table 7. Comparison of network experimental effects based on RDAB

      View table
      View in Article

      Table 7. Comparison of network experimental effects based on RDAB

      Fused operatorIntegrated RDAB
      Performance effect/timesNetwork size/times
      Focus1.281.06
      SPP+ Focus1.351.08
    Tools

    Get Citation

    Copy Citation Text

    Renhao Cai, Ning Cheng, Zhiyong Peng, Shize Dong, Jianmin An, Gang Jin. Lightweight infrared dim vehicle target detection algorithm based on deep learning[J]. Infrared and Laser Engineering, 2022, 51(12): 20220253

    Download Citation

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

    Category: Image processing

    Received: Mar. 20, 2022

    Accepted: --

    Published Online: Jan. 10, 2023

    The Author Email:

    DOI:10.3788/IRLA20220253

    Topics