• Infrared and Laser Engineering
  • Vol. 52, Issue 9, 20220876 (2023)
Siyu Wang, Xiaogang Yang, Ruitao Lu, Qingge Li..., Jiwei Fan and Zhengjie Zhu|Show fewer author(s)
Author Affiliations
  • Missile Engineering Institute, PLA Rocket Force University of Engineering, Xi'an 710025, China
  • show less
    DOI: 10.3788/IRLA20220876 Cite this Article
    Siyu Wang, Xiaogang Yang, Ruitao Lu, Qingge Li, Jiwei Fan, Zhengjie Zhu. Infrared time-sensitive target detection technology based on cross-modal data augmentation[J]. Infrared and Laser Engineering, 2023, 52(9): 20220876 Copy Citation Text show less
    Overview of two-stage model for IR time-sensitive target data augmentation
    Fig. 1. Overview of two-stage model for IR time-sensitive target data augmentation
    Single-scale generative model structure
    Fig. 2. Single-scale generative model structure
    Overall structure of Yolov5s target detection
    Fig. 3. Overall structure of Yolov5s target detection
    The CSP-A structure of adding attention mechanism
    Fig. 4. The CSP-A structure of adding attention mechanism
    Modality transfer partial training dataset
    Fig. 5. Modality transfer partial training dataset
    Visible infrared image conversion results
    Fig. 6. Visible infrared image conversion results
    Partial generated image data
    Fig. 7. Partial generated image data
    Comparison of typical IR time-sensitive target detection results
    Fig. 8. Comparison of typical IR time-sensitive target detection results
    Test results on partial image
    Fig. 9. Test results on partial image
    MeanStandard deviationVarianceInformation entropyContrast ratioMean gradient
    Original IR images126.049 3143.281 210.029 377.231 47147.147 565.135 38
    Transfer IR images137.884 1545.599 760.032 647.188 5443.930 823.158 52
    Table 1. Comparison of modal transformation effects
    DatasetPrecisionRecallmAP@0.5F1
    Origin0.786 30.910 50.892 40.843 9
    Augmentation0.932 00.970 40.980 60.950 8
    Table 2. Performance comparison before and after data augmentation
    MethodPrecision rateRecall ratemAP@0.5mAP@0.5 (ship) mAP@0.5 (aircraft) F1
    SSD0.35640.84230.82710.76930.88480.5009
    Fast R-CNN0.43280.85340.83270.85640.81800.5743
    Yolov50.85840.91610.95320.96870.93760.8863
    Yolov5 (CSP-A)0.93200.97040.98060.98070.98050.9508
    Table 3. Comparison experiments of different detection methods
    NumberSECBAMmAP@0.5
    1--0.937 6
    2-0.956 8
    3-0.962 4
    40.980 5
    Table 4. Ablation experiment results
    Siyu Wang, Xiaogang Yang, Ruitao Lu, Qingge Li, Jiwei Fan, Zhengjie Zhu. Infrared time-sensitive target detection technology based on cross-modal data augmentation[J]. Infrared and Laser Engineering, 2023, 52(9): 20220876
    Download Citation