Journal of Infrared and Millimeter Waves, Volume. 44, Issue 2, 285(2025)

A lightweight dark object detection network for infrared images

Zhao-Xu LI1... Qing-Xu XU1, Wei AN1,*, Xu HE1, Gao-Wei GUO1, Miao LI1,**, Qiang LING1, Long-Guang WANG2, Chao XIAO1 and Zai-Ping LIN1 |Show fewer author(s)
Author Affiliations
  • 1College of Electronic Science and Technology,National University of Defense Technology,Changsha 410073,China
  • 2Aviation University of Air Force,Changchun 130000,China
  • show less
    Figures & Tables(14)
    The thermal infrared images of real civial airplanes capured by SDGSAT-1:(a)8~10.5 μm;(b)10.3~11.3 μm;(c)11.5~12.5 μm
    Schematic diagram of AirFormer network structure
    Schematic diagram of the calculation for object abundance matrix: (a) object shape modeling; (b) shape model embedding in image; (c) object abundance matrix calculation; (d) Gaussian blurring of the abundance matrix
    The 1st real civial aircraft and its simulation: (a) simulated image; (b) real object; (c) simulated object
    The 5th real civial aircraft and its simulation: (a) simulated image; (b) real object; (c) simulated object
    Simulated sequence examples: (a) sequence 0041; (b) sequence 0077; (c) sequence 0266; (d) sequence 0393
    The detection results of AirFormer for real civil airports: (a) the 1st real airport; (b) the 2nd real airport; (c) the 3rd real airport; (d) the 4th real airport; (e) the 5th real airport
    • Table 1. The imaging information of real civial airplanes

      View table
      View in Article

      Table 1. The imaging information of real civial airplanes

      序号经度纬度日期当地时间

      局部

      背景

      目标一125.22° E30.86°N23.03.1609:37
      目标二115.16°E39.65°N23.03.2220:30陆地、云
      目标三123.71°E37.32°N23.08.0309:42
      目标四118.42°E34.28°N23.10.0109:01陆地、云
      目标五126.34°E37.21°N23.10.1720:41
    • Table 2. Grayscale value information of real civial airplanes and background

      View table
      View in Article

      Table 2. Grayscale value information of real civial airplanes and background

      序号波段

      目标

      灰度值

      局部背景

      平均灰度值

      差值比例
      目标一B1141915629.17%
      B2164418179.53%
      B3117712606.59%
      目标二B1101410442.92%
      B2119812393.34%
      B38588751.98%
      目标三B1180419175.91%
      B2203621645.96%
      B3139814614.37%
      目标四B1165217244.21%
      B2188019674.44%
      B3131813593.08%
      目标五B1166817735.95%
      B2192820415.58%
      B3137014253.88%
    • Table 3. Simulation parameter settings for real objects

      View table
      View in Article

      Table 3. Simulation parameter settings for real objects

      参数目标1目标2
      坐标(10.2,10.7)(10.3,9.95)
      目标长度80m80m
      航向角45°
      差值比例0.180.11
      高斯模糊标准差0.70.7
    • Table 4. Sample numbers under different load parameters on the IRAir dataset

      View table
      View in Article

      Table 4. Sample numbers under different load parameters on the IRAir dataset

      场景

      类型

      训练集

      序列数

      训练集目标数测试集序列数测试集目标数
      波段B130119703382260
      B236322713452332
      B333621313171977
      帧频1~5 FPS48030594993229
      6~10 FPS52033135013340

      帧间

      位移

      0像素33921583152063
      1像素33921573612374
      2像素32220573242132

      噪声

      强度

      0.00249732004673079
      0.00550331725333490
      总计1000637210006569
    • Table 5. The parameter settings of example simulated sequences

      View table
      View in Article

      Table 5. The parameter settings of example simulated sequences

      参数

      序列

      0041

      序列

      0077

      序列

      0266

      序列

      0393

      波段B2B2B3B3
      帧频1 FPS6 FPS2 FPS10 FPS
      帧间位移0像素1像素2像素2像素
      噪声强度0.0020.0020.0050.002
      目标数量41085
    • Table 6. Performance comparison of detection methods

      View table
      View in Article

      Table 6. Performance comparison of detection methods

      方法CornerNetYOLOv3Deformable DETRRTMDET-tinyYOLOX-tinyDSFNetAirFormer
      AP0.3360.2700.2740.3500.3980.2330.349
      AP200.7700.7520.7090.8120.7650.5280.737
      召回率0.7380.7660.6880.5440.7160.5040.710
      准确率0.9040.9020.8970.6750.8430.9320.826
      F10.8120.8280.7790.6030.7740.6530.764
      参数量201.0M61.5M41.1M2.7M2.7M17.0M37.1K
      FLOPs112.8G12.4G15.0G5.9G5.5G12.2G46.2M
      推理耗时29.4 ms11.7 ms32.3 ms10.6 ms9.1 ms50.1 ms5.7 ms
    • Table 7. Detection performance comparison of AirFormer for objects with different sizes

      View table
      View in Article

      Table 7. Detection performance comparison of AirFormer for objects with different sizes

      目标长度/m真值数检出数召回率/%
      406 2223 01648.5
      506 1544 62675.2
      602 1091 83887.2
      702 0161 79889.2
      802 0621 90992.6
    Tools

    Get Citation

    Copy Citation Text

    Zhao-Xu LI, Qing-Xu XU, Wei AN, Xu HE, Gao-Wei GUO, Miao LI, Qiang LING, Long-Guang WANG, Chao XIAO, Zai-Ping LIN. A lightweight dark object detection network for infrared images[J]. Journal of Infrared and Millimeter Waves, 2025, 44(2): 285

    Download Citation

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

    Category: Interdisciplinary Research on Infrared Science

    Received: May. 8, 2024

    Accepted: --

    Published Online: Mar. 14, 2025

    The Author Email: AN Wei (anwei@nudt.edu.cn), LI Miao (lm8866@nudt.edu.cn)

    DOI:10.11972/j.issn.1001-9014.2025.02.016

    Topics