Deep learning based measurement accuracy improvement of high dynamic range objects in fringe projection profilometry

Area of Science:

• Optics and Photonics
• Computer Vision
• Metrology

Background:

• Phase retrieval accuracy is crucial for 3D measurement in fringe projection profilometry (FPP).
• High dynamic range (HDR) objects present challenges like fringe saturation and low contrast in FPP.
• Traditional FPP methods require numerous images for HDR objects, limiting applications with moving objects.

Purpose of the Study:

• To propose a novel deep learning-based method for accurate phase demodulation in FPP of HDR objects.
• To address the limitations of traditional FPP methods in capturing HDR object details.
• To enable faster and more reliable 3D measurements of HDR objects, including those in motion.

Main Methods:

• An improved UNet deep neural network was utilized to establish a 'many-to-one' mapping for phase demodulation.
• π-shifted binary fringes were employed to capture more saturated fringe information.
• The deep learning model was trained to directly retrieve wrapped phases from fringe patterns.

Main Results:

• The proposed deep learning method successfully retrieves wrapped phases for HDR objects with high accuracy.
• The method overcomes challenges of fringe saturation and low contrast inherent in HDR object measurement.
• Experimental results validated the effectiveness and reliability of the deep learning approach.

Conclusions:

• Deep learning, specifically an improved UNet with π-shifted binary fringes, offers a robust solution for phase demodulation in FPP of HDR objects.
• This method significantly improves 3D measurement accuracy and speed for challenging HDR targets.
• The findings pave the way for advanced FPP applications in dynamic and high-contrast environments.