• Laser & Optoelectronics Progress
  • Vol. 60, Issue 12, 1215003 (2023)
Zongbao Bai1, Junju Zhang1,*, Yuan Gao2, and Youcheng Hu1
Author Affiliations
  • 1School of Electronic and Optical Engineering, Nanjing University of Science & Technology, Nanjing 210094, Jiangsu, China
  • 2School of Electronic and Optical Engineering, Nanjing University of Science and Technology ZiJin College, Nanjing 210023, Jiangsu, China
  • show less
    DOI: 10.3788/LOP221025 Cite this Article Set citation alerts
    Zongbao Bai, Junju Zhang, Yuan Gao, Youcheng Hu. Attention Mechanism-Based Object Detection Algorithm in Aerial Images[J]. Laser & Optoelectronics Progress, 2023, 60(12): 1215003 Copy Citation Text show less
    Structure of Faster R-CNN
    Fig. 1. Structure of Faster R-CNN
    Structure diagram of Tri-CSAM
    Fig. 2. Structure diagram of Tri-CSAM
    Channel feature maps of ResNet-50 in different nodes. (a) Original image; (b) convolution layer of Conv1;(c)rectified linear unit of Conv1;(d)output layer of Conv2_x
    Fig. 3. Channel feature maps of ResNet-50 in different nodes. (a) Original image; (b) convolution layer of Conv1;(c)rectified linear unit of Conv1;(d)output layer of Conv2_x
    Schematic of multi-scale receptive field
    Fig. 4. Schematic of multi-scale receptive field
    Structure of DH
    Fig. 5. Structure of DH
    Structure of regression branch convolution module. (a) Structure of block 1; (b) structure of block 2
    Fig. 6. Structure of regression branch convolution module. (a) Structure of block 1; (b) structure of block 2
    Label visualization of partial dataset
    Fig. 7. Label visualization of partial dataset
    P-R curves under different IoU thresholds. (a) IoU is 0.50; (b) IoU is 0.75; (c) IoU is 0.90
    Fig. 8. P-R curves under different IoU thresholds. (a) IoU is 0.50; (b) IoU is 0.75; (c) IoU is 0.90
    Visualization results of network heat map. (a) Small target pedestrian scene of 45° aerial view; (b) medium target vehicles scene of 45° aerial view; (c) small target vehicles scene of 90° aerial view
    Fig. 9. Visualization results of network heat map. (a) Small target pedestrian scene of 45° aerial view; (b) medium target vehicles scene of 45° aerial view; (c) small target vehicles scene of 90° aerial view
    Visualization results of aerial image detection. (a) Normal light scene; (b) week light scene; (c) strong light scene
    Fig. 10. Visualization results of aerial image detection. (a) Normal light scene; (b) week light scene; (c) strong light scene
    ModelTri-CSAMDHAP50AP75APSAPMAPLmAP
    Baseline40.1223.529.2335.6644.4223.22
    G141.1624.8110.6136.4844.8924.29
    G242.0325.2710.3336.7145.2524.35
    Proposed model42.9125.7411.2536.8244.9725.63
    Table 1. Detection results on VisDrone dataset
    ModelTri-CSAMDHAP50AP75APSAPMAPLmAP
    Baseline42.3126.6110.0835.7848.6924.65
    G145.8729.7912.5840.3650.9327.76
    G244.9129.2611.9839.8851.3627.02
    Proposed model46.3930.0212.4642.2651.7329.16
    Table 2. Detection results on UAVDT dataset
    ModelmAP /%AP50 /%AP75 /%Speed /(frame·s-1
    Faster R-CNN23.2240.1223.5219
    Mask R-CNN24.1641.1623.8711
    Cascade R-CNN16.0931.9115.0113
    YOLOv319.8338.2320.6330
    DPN25.0950.0121.83
    TridentNet22.5143.2920.50
    CornerNet17.4134.1215.7816
    LResNet22.3239.6323.17
    Proposed model25.6342.9125.7417
    Table 3. Performance comparison of different algorithms
    Zongbao Bai, Junju Zhang, Yuan Gao, Youcheng Hu. Attention Mechanism-Based Object Detection Algorithm in Aerial Images[J]. Laser & Optoelectronics Progress, 2023, 60(12): 1215003
    Download Citation