• Laser & Optoelectronics Progress
  • Vol. 58, Issue 20, 2010013 (2021)
Yanlei Xu1, Jiran Liang1、2、*, Guojun Dong3, and Zhuang Chen1
Author Affiliations
  • 1School of Microelectronics, Tianjin University, Tianjin 300072, China
  • 2Tianjin Key Laboratory of Imaging and Sensing Microelectronic Technology, Tianjin 300072, China
  • 3Tianjin 712 Communication & Broadcasting Shareholding Co., Ltd., Tianjin 300457, China;
  • show less
    DOI: 10.3788/LOP202158.2010013 Cite this Article Set citation alerts
    Yanlei Xu, Jiran Liang, Guojun Dong, Zhuang Chen. Aerial Image Target Detection Algorithm Based on Improved CenterNet[J]. Laser & Optoelectronics Progress, 2021, 58(20): 2010013 Copy Citation Text show less
    Algorithm flowchart
    Fig. 1. Algorithm flowchart
    Schematic diagram of deformable cavity convolution structure
    Fig. 2. Schematic diagram of deformable cavity convolution structure
    CBA-connection. (a) Overall structure diagram; (b) attention module flowchart
    Fig. 3. CBA-connection. (a) Overall structure diagram; (b) attention module flowchart
    Schematic diagram of backbone network structure
    Fig. 4. Schematic diagram of backbone network structure
    Loss graph of different models
    Fig. 5. Loss graph of different models
    Aerial image target detection effect display. (a) Structured information dropout; (b) false and missed detection targets construct new samples; (c) night scene; (d) strong light scene; (e) (f) gathering area
    Fig. 6. Aerial image target detection effect display. (a) Structured information dropout; (b) false and missed detection targets construct new samples; (c) night scene; (d) strong light scene; (e) (f) gathering area
    BackbonemAP /%AP50 /%AP75 /%FPS /(frame·s-1)
    ResNet-5020.0542.0519.7565
    DLA-3422.5045.8820.5055
    ResDcn-1814.0136.0015.25131
    Ours25.2245.6223.3645
    Table 1. Comparison of detection effects of different backbone networks
    BackbonemAP /%AP50 /%AP75 /%
    ResNet-5020.0542.0519.75
    Ours+ (D-ASPP)22.5743.6320.97
    Ours++ (CBA-connection)24.4145.7122.13
    Ours+++ (Threshold)25.2245.6223.36
    Table 2. Comparison of effectiveness after successively introducing the designed structure
    Target typeAP /%Target typeAP /%
    Airplane29.77Basketball court17.69
    Ship26.94Ground track field30.64
    Storage tank32.32Harbor29.31
    Baseball diamond31.29Bridge17.16
    Tennis court16.18Vehicle20.90
    Table 3. Evaluation of the recognition accuracies of different target types in the NWPU VHR-10 data set
    MethodBackbonemAP /%AP50 /%AP75 /%FPS /(frame·s-1)
    Faster R-CNNResNet-10127.2348.5724.509
    YOLOv3DarkNet-5322.4945.2219.5041
    RetinaNetResNet-5016.8830.2616.0519
    CornerNetHourglass-10419.1439.9017.7922
    OursResNet+25.2245.6223.3645
    Table 4. Comparison of the detection effects of different detection algorithms on the NWPU VHR-10 data set
    Yanlei Xu, Jiran Liang, Guojun Dong, Zhuang Chen. Aerial Image Target Detection Algorithm Based on Improved CenterNet[J]. Laser & Optoelectronics Progress, 2021, 58(20): 2010013
    Download Citation