• Acta Optica Sinica
  • Vol. 41, Issue 23, 2311001 (2021)
Bofan Wang and Haitao Zhao*
Author Affiliations
  • School of Information Science and Engineering, East China University of Science and Technology, Shanghai 200237, China
  • show less
    DOI: 10.3788/AOS202141.2311001 Cite this Article Set citation alerts
    Bofan Wang, Haitao Zhao. Small Object Detection in Hyperspectral Images Based on Radial Basis Activation Function[J]. Acta Optica Sinica, 2021, 41(23): 2311001 Copy Citation Text show less
    Schematic diagram of the radial basis activation function for spectral screening (RBAF-SS)
    Fig. 1. Schematic diagram of the radial basis activation function for spectral screening (RBAF-SS)
    Attention-based resolution reconstruction network (ABRRN)
    Fig. 2. Attention-based resolution reconstruction network (ABRRN)
    Radial basis object output network (RBOON)
    Fig. 3. Radial basis object output network (RBOON)
    Examples of hyperspectral data sets. (a) Small objects; (b) medium objects
    Fig. 4. Examples of hyperspectral data sets. (a) Small objects; (b) medium objects
    Qualitative analysis of spectral selection based on RBAF
    Fig. 5. Qualitative analysis of spectral selection based on RBAF
    Spectral curve of airplane object features
    Fig. 6. Spectral curve of airplane object features
    Experimental results of the proposed method are compared with four approaches with sigmoid based object output layers. (a) Faster RCNN[1] (ResNet-50); (b) YOLOv3[25] (Darknet-53); (c) FCOS[32] (ResNet-50); (d) CenterNet[24] (ResNet-18); (e) proposed method (ResNet-18)
    Fig. 7. Experimental results of the proposed method are compared with four approaches with sigmoid based object output layers. (a) Faster RCNN[1] (ResNet-50); (b) YOLOv3[25] (Darknet-53); (c) FCOS[32] (ResNet-50); (d) CenterNet[24] (ResNet-18); (e) proposed method (ResNet-18)
    False alarm rate of different approaches under different IoU threshold
    Fig. 8. False alarm rate of different approaches under different IoU threshold
    StageSub-networkBlockLayer detailsOutput shape /(pixel×pixel×pixel)
    Input----384×192×25
    1ABRRN-DeconvConv+ReLUConvRBAF-SSGAP768×384×6
    ABRRN-Element-wise multiplicationSame as the above ABRRN1536×768×3
    2ResNet-18backboneDown samplingConv+BN+ReLU384×192×64
    Max pooling
    ResBlock(no down sampling)Conv+BN+ReLUConv+BNIdentity384×192×64
    Element-wise Addition
    ResBlock(down sampling)Conv+BN+ReLUConv+BNConv+BN192×96×128
    Element-wise addition
    ResBlock(no down sampling)Same as the above ResBlock(no down sampling)192×96×128
    ResBlock(down sampling)Same as the above ResBlock(down sampling)96×48×256
    ResBlock(no down sampling)Same as the above ResBlock(no down sampling)96×48×256
    ResBlock(down sampling)Same as the above ResBlock(down sampling)48×24×512
    Up sampling blockDCN+BN+ReLUDeconv+BN+ReLU96×48×256
    3Up samplingnetworkUp sampling blockSame as the above upsampling block96×48×128
    Up sampling blockSame as the above upsampling block96×48×64
    4RBOON-Conv+ReLUConv+ReLUConv+ReLU
    RBAF-SODConvConv96×48×1/2/2
    Table 1. Overall structure of the detection network
    MethodsAP50 /%AP50:95 /%Time /s
    SmallMediumAllSmallMediumAll
    Faster RCNN[1] (ResNet-50)59.469.562.221.830.424.20.044
    YOLOv3[25](Darknet-53)51.867.456.218.528.521.30.015
    FCOS[32](ResNet-50)25.046.731.19.09.78.60.041
    Table 2. Detection accuracy and ablation experiment
    MethodsAP50 /%AP50:95 /%Time /s
    SmallMediumAllSmallMediumAll
    CenterNet[24](ResNet-18)60.270.463.121.528.322.70.011
    Proposed-ABRRN(ResNet-18)64.973.867.423.427.024.40.020
    Ours-RBOON(ResNet-18)60.770.663.521.628.422.90.011
    Proposed(ResNet-18)65.074.267.623.526.724.40.020
    Table 3. Ablation experiment
    Bofan Wang, Haitao Zhao. Small Object Detection in Hyperspectral Images Based on Radial Basis Activation Function[J]. Acta Optica Sinica, 2021, 41(23): 2311001
    Download Citation