• Acta Optica Sinica
  • Vol. 42, Issue 9, 0915001 (2022)
Rongchang Wang1、2, Feng Wang1、2、*, Shuaijun Ren1、2, and Yong Wang1、2
Author Affiliations
  • 1Department of Information Engineering, PLA Army Artillery Air Defense Force College, Hefei 230031, Anhui, China
  • 2Key Laboratory of Polarized Light Imaging Detection Technology of Anhui Province, Hefei 230031, Anhui, China
  • show less
    DOI: 10.3788/AOS202242.0915001 Cite this Article Set citation alerts
    Rongchang Wang, Feng Wang, Shuaijun Ren, Yong Wang. Polarization Imaging Detection of Individual Camouflage Based on Two-Stream Fusion Network[J]. Acta Optica Sinica, 2022, 42(9): 0915001 Copy Citation Text show less
    Schematic of light propagation
    Fig. 1. Schematic of light propagation
    Layout diagram of color focal plane polarized pixel array
    Fig. 2. Layout diagram of color focal plane polarized pixel array
    Structure diagram of TSF-Net
    Fig. 3. Structure diagram of TSF-Net
    Structure diagram of ANN
    Fig. 4. Structure diagram of ANN
    Structure diagram of APP-Net
    Fig. 5. Structure diagram of APP-Net
    Structure diagram of APP-Net
    Fig. 6. Structure diagram of APP-Net
    Structure diagram of RGB-Net
    Fig. 7. Structure diagram of RGB-Net
    Process of feature extraction and feature fusion
    Fig. 8. Process of feature extraction and feature fusion
    Schematic of training and test process
    Fig. 9. Schematic of training and test process
    Physical drawing of portable acquisition equipment
    Fig. 10. Physical drawing of portable acquisition equipment
    Schematic of classification of individual camouflage polarization image dataset
    Fig. 11. Schematic of classification of individual camouflage polarization image dataset
    Two types of camouflage target test diagram. (a) Multicam type camouflage; (b) Woodland type camouflage
    Fig. 12. Two types of camouflage target test diagram. (a) Multicam type camouflage; (b) Woodland type camouflage
    Detection effects of different models in Multicam dataset. (a) SSD model; (b) YOLOv4 model; (c) YOLOv5 model; (d) RetinaNet model; (e) Faster R-CNN model; (f) TSF-Net model
    Fig. 13. Detection effects of different models in Multicam dataset. (a) SSD model; (b) YOLOv4 model; (c) YOLOv5 model; (d) RetinaNet model; (e) Faster R-CNN model; (f) TSF-Net model
    Detection effects of different models in Woodland dataset. (a) SSD model; (b) YOLOv4 model; (c) YOLOv5 model; (d) RetinaNet model; (e) Faster R-CNN model; (f) TSF-Net model
    Fig. 14. Detection effects of different models in Woodland dataset. (a) SSD model; (b) YOLOv4 model; (c) YOLOv5 model; (d) RetinaNet model; (e) Faster R-CNN model; (f) TSF-Net model
    Parameter verification result
    Fig. 15. Parameter verification result
    Verification results for different branches. (a) Detection accuracy of different v values; (b) IOU-mAP curves
    Fig. 16. Verification results for different branches. (a) Detection accuracy of different v values; (b) IOU-mAP curves
    StructureGPU memoryusage /MBTime/minLoss
    (8,16,8,3)14532251.23×10-2
    (16,8,8,3)14532169.51×10-3
    (96,48,32,3)38172077.48×10-3
    (48,96,32,3)38172295.79×10-3
    (128,96,64,32,3)61092555.55×10-3
    (96,128,64,32,3)61092824.27×10-3
    Table 1. Training results with different structures
    CategoryParameter
    Camera modelFLIR BFS-U3-51S5PC-C
    Resolution /(pixel×pixel)2448×2048
    Frame rate /(frame·s-1)75
    Chip modelSony IMX250MYR,Polar-RGB
    Data interfaceUSB3.1 Gen1
    Size and weight /(mm×mm×mm)29×29×30
    Mass /g36
    Lens interfaceC-Mount
    Table 2. Parameters of color focal plane camera
    CasePrediction (positive)Prediction (negative)
    Ture(true)TPTN
    Ture(false)FPFN
    Table 3. Positive and negative cases
    ModelmAP /%
    Multicam datesetWoodland dataset
    SSD70.973.5
    YOLOv471.573.4
    YOLOv573.174.6
    RetinaNet75.277.5
    Faster R-CNN77.178.9
    TSF-Net85.987.1
    Table 4. Comparison of detection accuracy of different models
    ModelmAP /%
    SFSSSB
    TSF-Net84.984.384.6
    Table 5. Comparison of posture test results of different camouflage personnel
    ModelmAP /%
    M/WW/M
    Faster R-CNN23.515.7
    TSF-Net48.835.5
    Table 6. Cross-validation comparison
    Rongchang Wang, Feng Wang, Shuaijun Ren, Yong Wang. Polarization Imaging Detection of Individual Camouflage Based on Two-Stream Fusion Network[J]. Acta Optica Sinica, 2022, 42(9): 0915001
    Download Citation