• Acta Photonica Sinica
  • Vol. 51, Issue 4, 0410003 (2022)
Haicheng QU and Lei SHEN*
Author Affiliations
  • College of Software,Liaoning Technical University,Huludao,Liaoning 125105,China
  • show less
    DOI: 10.3788/gzxb20225104.0410003 Cite this Article
    Haicheng QU, Lei SHEN. SAR Image Denoising and Semantic Enhancement for Object Detection[J]. Acta Photonica Sinica, 2022, 51(4): 0410003 Copy Citation Text show less
    Network structure diagram
    Fig. 1. Network structure diagram
    Receptive area mapping
    Fig. 2. Receptive area mapping
    Pixel-level denoising diagram
    Fig. 3. Pixel-level denoising diagram
    Pixel-level denoising effect
    Fig. 4. Pixel-level denoising effect
    Semantic enhanced structure
    Fig. 5. Semantic enhanced structure
    Asymmetric convolutional layer structure
    Fig. 6. Asymmetric convolutional layer structure
    Structure of Transformer Encoder
    Fig. 7. Structure of Transformer Encoder
    Different backbones network detection results
    Fig. 8. Different backbones network detection results
    Different modules detection effect
    Fig. 9. Different modules detection effect
    P-R diagram of different comparison methods
    Fig. 10. P-R diagram of different comparison methods
    Different methods detection effect
    Fig. 11. Different methods detection effect
    Different scenes detection effect diagram
    Fig. 12. Different scenes detection effect diagram
    DatasetsSensorsResolution/mPolarizationScenes
    SSDD

    RadarSat-2

    TerraSAR-X

    Sentinel-1

    1~5HH,HV,VV,VH

    Offshore,Inshore,

    Coherent speckle noise

    SAR-Ship-Dataset 17

    Sentinel-1

    GF-3

    3,5,8,10,25HH,HV,VV,VH
    HRSID 18

    Sentinel-1B

    TerraSAR-X

    TanDEM

    0.5,1,3HH,HV,VV,VH
    Table 1. Data set parameters
    BackboneRecall/%Precision/%Average precision/%
    Resnet5097.0795.1396.53
    Resnet10197.5896.1196.73
    Table 2. Detection performance of different backbone networks
    Pixel-level noise reductionSemantic enhancementRecall/%Precision/%Average precision/%
    ××89.9494.2689.23
    ×96.9496.3396.44
    ×96.9395.8496.40
    97.5896.1196.73
    Table 3. Detection performance of different models
    Pixel-level noise reductionSemantic enhancementParams/MTraining time /FPSTesting time /FPS
    ××152.766.8922.64
    ×160.846.4920.50
    ×151.406.1020.92
    159.495.9921.09
    Table 4. Model performance analysis
    MethodsBackboneRecall/%Precision/%Average precision/%
    Faster R-CNNResnet10188.5488.0987.26
    RetinanetResnet10189.1787.9486.91
    FPNResnet10192.2286.4190.69
    YOLOv4CSPDarknet53--95.60
    WANG24ShuffleNetV2--94.70
    R2FA-Det25Resnet50--94.72
    FBR-Net 26-92.7994.0194.10
    OursResnet10197.5896.1196.73
    Table 5. Detection performance of different methods on SSDD
    MethodsBackboneRecall/%Precision/%Average precision /%
    Faster R-CNNResnet10180.2481.3976.35
    RetinanetResnet10172.1790.6271.03
    FPNResnet10183.4090.8081.38
    OursResnet10191.8681.4588.86
    Table 6. Detection performance of different methods on SAR-Ship-Dataset
    ScenesMethodsRecall/%Precision/%Average precision/%
    InshoreBaseline79.0187.0777.79
    Our91.9889.7690.00
    OffshoreBaseline92.7896.0192.24
    Our99.0497.7898.53
    Small-scale targetBaseline90.7990.7890.66
    Our97.6096.7896.85
    Large-scale targetBaseline89.8494.6589.13
    Our97.3792.5094.64
    Table 7. Detection performance of different scenes
    Haicheng QU, Lei SHEN. SAR Image Denoising and Semantic Enhancement for Object Detection[J]. Acta Photonica Sinica, 2022, 51(4): 0410003
    Download Citation