Advanced Imaging, Volume. 1, Issue 2, 021004(2024)

Real-time 3D imaging based on ROI fringe projection and a lightweight phase-estimation network

Yueyang Li, Junfei Shen, Zhoujie Wu, Yajun Wang, and Qican Zhang*
Author Affiliations
  • College of Electronics and Information Engineering, Sichuan University, Chengdu, China
  • show less
    Figures & Tables(15)
    Flowchart of real-time 3D imaging based on ROI projection and phase estimation. The green and orange arrows indicate the processing flow for low- and high-frequency fringe patterns, respectively.
    Real-time phase estimation network structure. (a) Outer structure of PE-Net, loss function design, and training process; (b) internal structure of the phase estimation module; (c) internal structure of the lightweight network.
    NonBt1d module, downsampling module, dilated convolution module, and upsampling module within the lightweight network.
    Flowchart of the MHPU method. (a) Equivalent phase with the phase difference of low- and high-frequency fringe; (b) unwrapped equivalent phase assisted by a reference plane; (c) unwrapped phases of low frequency and high frequency.
    Depth limitation of MHPU.
    Two fringe projection strategies. (a) Normal fringe projection. (b) ROI fringe projection.
    Comparison of two fringe projection strategies. (a) Generated fringe pattern (top left) and perspective view (top right) of normal fringe projection. (b) Corresponding results of ROI fringe projection.
    Measurement results of standard spheres using three-step PSP and two projection strategies. (a), (b) Fringe patterns of two projection strategies. (c),(d) 3D reconstruction results of two projection strategies.
    Train and validation loss curves of PE-Net.
    3D reconstruction results of the David sculpture. (a) Ground truth. (b) Feng’s method; (c) UNet method; (d) NAS method; (e) proposed PE-Net.
    Error distributions of the David sculpture. (a) Feng’s method; (b) UNet method; (c) NAS method; (d) proposed PE-Net. (e)–(h) Localized enlarged images of the dashed-boxed regions corresponding to each method.
    3D reconstruction results and error values of the standard spheres. (a) Feng’s method; (c) UNet method; (e) NAS method; (g) proposed PE-Net. (b), (d), (f), (h) Corresponding depth error distributions and error values.
    Real-time 3D reconstruction results (Visualization 1). (a), (b) Measurement scene comprising a stationary David and a rotating Nike sculpture. (c)–(f) 3D point cloud of different frames.
    • Table 1. Quantitative Results of Standard Sphere Measurement Using Three-Step PSP and Two Projection Strategies

      View table
      View in Article

      Table 1. Quantitative Results of Standard Sphere Measurement Using Three-Step PSP and Two Projection Strategies

       DADBDCMAEAMAEBRMSARMSB
      Normal50.800250.8158100.11400.09950.08550.12070.1100
      ROI50.800050.8062100.13620.07700.07180.09270.0857
    • Table 2. MAE and RMS of the Phase Error, Inference Time, and Model Size of Four Methods

      View table
      View in Article

      Table 2. MAE and RMS of the Phase Error, Inference Time, and Model Size of Four Methods

      MethodMAE (rad)RMS (rad)Inference time (ms)Model size (M)
      Feng0.034170.0581167.700.70
      UNet0.031030.0500741.0231.00
      NAS0.037280.0636417.020.13
      PE-Net0.029790.051636.012.10
    Tools

    Get Citation

    Copy Citation Text

    Yueyang Li, Junfei Shen, Zhoujie Wu, Yajun Wang, Qican Zhang, "Real-time 3D imaging based on ROI fringe projection and a lightweight phase-estimation network," Adv. Imaging 1, 021004 (2024)

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category: Research Article

    Received: Jun. 11, 2024

    Accepted: Sep. 2, 2024

    Published Online: Sep. 25, 2024

    The Author Email: Qican Zhang (zqc@scu.edu.cn)

    DOI:10.3788/AI.2024.10008

    Topics