Deep learning-based high dynamic range 3D reconstruction.

Area of Science:

• Optics and Photonics
• Computer Vision
• Machine Learning

Background:

• Three-dimensional (3D) reconstruction using fringe projection profilometry (FPP) is vital for industrial manufacturing.
• Overexposure in FPP images due to varying reflectance and lighting degrades 3D reconstruction accuracy, especially in high dynamic range (HDR) environments.

Purpose of the Study:

• To develop a deep learning-based method for restoring saturated fringe images in HDR scenes.
• To enhance the accuracy of 3D reconstruction by addressing overexposure issues in FPP.

Main Methods:

• A novel deep learning approach utilizing U-Net derivative networks for fringe image restoration.
• Systematic comparison of U-Net, Res-U-Net, and SE-U-Net architectures for fringe repair.
• Quantitative experimental analysis to evaluate network performance.

Main Results:

• All tested deep learning networks (U-Net, Res-U-Net, SE-U-Net) effectively repaired saturated fringe images.
• SE-U-Net showed superior performance in restoring missing image regions.
• The proposed method significantly improves 3D reconstruction accuracy without extra hardware.

Conclusions:

• Deep learning is effective for restoring saturated fringe images in HDR scenes.
• The study provides guidance on selecting appropriate network models for fringe restoration tasks.
• This method offers a practical solution for improving 3D reconstruction in challenging lighting conditions.