• Infrared and Laser Engineering
  • Vol. 52, Issue 10, 20230010 (2023)
Guang Zhai, Shengran Hu, and Yiyong Sun*
Author Affiliations
  • School of Aerospace Engineering, Beijing Institute of Technology, Beijing 100081, China
  • show less
    DOI: 10.3788/IRLA20230010 Cite this Article
    Guang Zhai, Shengran Hu, Yiyong Sun. Research on LSTM method of high dynamic dim and small targets detection for space-based infrared early warning[J]. Infrared and Laser Engineering, 2023, 52(10): 20230010 Copy Citation Text show less
    Structure of LSTM network unit
    Fig. 1. Structure of LSTM network unit
    Process of LISTD algorithm
    Fig. 2. Process of LISTD algorithm
    Preprocessing block. \begin{document}$(\rm a)$\end{document} Block process; \begin{document}$(\rm b)$\end{document} Filter core
    Fig. 3. Preprocessing block. \begin{document}$(\rm a)$\end{document} Block process; \begin{document}$(\rm b)$\end{document} Filter core
    LSTM detection block
    Fig. 4. LSTM detection block
    Simulation image
    Fig. 5. Simulation image
    LISTD detection results of sinusoidal moving target
    Fig. 6. LISTD detection results of sinusoidal moving target
    Detection result. \begin{document}$(\rm a)$\end{document} False alarm scene; \begin{document}$(\rm b)$\end{document} Mistaken detection scene; \begin{document}$(\rm c)$\end{document} Missing alarm scene; \begin{document}$(\rm d)$\end{document} Correct detection scene
    Fig. 7. Detection result. \begin{document}$(\rm a)$\end{document} False alarm scene; \begin{document}$(\rm b)$\end{document} Mistaken detection scene; \begin{document}$(\rm c)$\end{document} Missing alarm scene; \begin{document}$(\rm d)$\end{document} Correct detection scene
    Detection result: PositiveDetection result: Negative
    Reality: PositiveTPFN
    Reality: NegativeFPTN
    Table 1. Possible situations in target detection
    Uniform linear motionSinusoidal motionComplex motionAverage time consumption/帧·s−1Peak memory usage/MiBNeural network GFLOPs
    PrecisionRecallPrecisionRecallPrecisionRecall
    MF0.64440.83110.63870.83430.64150.83740.007253360.44-
    MMF0.45130.87150.44930.86770.45100.87032.7241359.24-
    SHT0.99540.74040.96300.70930.95850.68040.009351359.25-
    YOLOv50.36600.09490.37200.10100.34000.09240.06760597.3916.422
    LISTD0.94830.92370.93470.88510.93620.86330.0083261214.130.108
    Table 2. Performance comparison among algorithms
    SCRSNRUniform linear motionSinusoidal motionComplex motion
    PrecisionRecallPrecisionRecallPrecisionRecall
    220.69050.34180.73220.35490.65420.3777
    40.81910.60140.79740.61680.84330.6145
    60.86740.62690.82860.64300.84150.6547
    420.84600.61310.88420.66990.87090.6200
    40.87230.74980.89580.77990.87400.7264
    60.90950.78450.93990.80820.85070.7657
    620.89710.74480.92290.75630.87780.6653
    40.92900.87520.95000.88650.92820.8577
    60.94870.88510.94540.92630.95570.9161
    Table 3. LISTD detection results for datasets with different SCR and SNR
    Guang Zhai, Shengran Hu, Yiyong Sun. Research on LSTM method of high dynamic dim and small targets detection for space-based infrared early warning[J]. Infrared and Laser Engineering, 2023, 52(10): 20230010
    Download Citation