• Acta Optica Sinica
  • Vol. 40, Issue 7, 0720001 (2020)
Changdong Yu1、**, Xiaojun Bi2、*, Yang Han3, Haiyun Li1, and Yunfei Gui3
Author Affiliations
  • 1College of Information and Communication Engineering, Harbin Engineering University,Harbin, Heilongjiang 150001, China
  • 2College of Information and Engineering, Minzu University of China, Beijing 100081, China
  • 3College of Shipbuilding Engineering, Harbin Engineering University, Harbin, Heilongjiang 150001, China
  • show less
    DOI: 10.3788/AOS202040.0720001 Cite this Article Set citation alerts
    Changdong Yu, Xiaojun Bi, Yang Han, Haiyun Li, Yunfei Gui. Particle Image Velocimetry Based on a Lightweight Deep Learning Model[J]. Acta Optica Sinica, 2020, 40(7): 0720001 Copy Citation Text show less
    Structure of the LiteFlowNet
    Fig. 1. Structure of the LiteFlowNet
    Improved NetE structure
    Fig. 2. Improved NetE structure
    Depth separable convolution
    Fig. 3. Depth separable convolution
    Velocity fields and vorticity maps of DNS turbulent flow estimated by different models
    Fig. 4. Velocity fields and vorticity maps of DNS turbulent flow estimated by different models
    RMSE estimated by different models for turbulent image
    Fig. 5. RMSE estimated by different models for turbulent image
    Histogram comparison of velocity distributions estimated by different models
    Fig. 6. Histogram comparison of velocity distributions estimated by different models
    Number of parameters for different models
    Fig. 7. Number of parameters for different models
    Layer nameKernelStrideRepeat timesOutput resolution
    conv132132, 256, 256
    conv2_132132, 128, 128
    conv2_231132, 128, 128
    conv2_331132, 128, 128
    conv3_1/dw32132, 64, 64
    conv3_1/sep12164, 64, 64
    conv3_2/dw31164, 64, 64
    conv3_2/sep11164, 64, 64
    conv4_1/dw32164, 32, 32
    conv4_1/sep12196, 32, 32
    conv4_2/dw32196, 32, 32
    conv4_2/sep12196, 32, 32
    conv5/dw32196, 16, 16
    conv5/sep121128, 16, 16
    Table 1. Improved NetC network structure parameters
    Case nameDescriptionQuantity
    UniformUniform flow1000
    Back-stepBackward stepping flow3200
    CylinderVortex shedding flow over a circular cylinder2050
    DNS-turbulenceAhomogeneous and isotropic turbulence flow2000
    SQGSeasurface flow driven by a Surface Quasi-Geostrophic model1500
    JHTDB-channelChannel flow provided by Johns Hopkins Turbulence Databases1600
    JHTIDB-mhd1024Forced MHD turbulence provided by JHTIDB800
    JHTIDB-isotropic1024Forced isotropic turbulence provided by JHTIDB2000
    Table 2. Types of motion fields included in the PIV dataset
    ModelRMSE
    TrainTest
    LiteFlowNet0.22500.2300
    LiteFlowNet-en0.07100.0730
    LiteFlowNet-HD0.06800.0682
    Table 3. Test errors on DNS
    ModelTime /msNumber of vectors
    LiteFlowNet24.74256×256
    LiteFlowNet-en46.54256×256
    LiteFlowNet-HD41.98256×256
    Table 4. Computation time of different models
    Changdong Yu, Xiaojun Bi, Yang Han, Haiyun Li, Yunfei Gui. Particle Image Velocimetry Based on a Lightweight Deep Learning Model[J]. Acta Optica Sinica, 2020, 40(7): 0720001
    Download Citation