FDCGAN: frequency domain constrained generative adversarial network for overexposed fringe image restoration

Area of Science:

• Computer Vision
• Metrology
• Image Processing

Background:

• Fringe projection profilometry (FPP) struggles with overexposed images, losing crucial structural details.
• Traditional restoration methods are often ineffective due to severe saturation in FPP images.

Purpose of the Study:

• To develop a robust method for restoring overexposed fringe images in FPP.
• To enhance the quality and reliability of 3D reconstructions from challenging FPP data.

Main Methods:

• Propose FDCGAN (Frequency Domain Constrained Generative Adversarial Network) combining spatial and frequency-domain learning.
• Utilize hierarchical feature encoding, multi-scale adversarial supervision, and normalization strategies.
• Incorporate frequency-aware losses (Fourier magnitude loss, high-frequency loss) for realistic fringe pattern reconstruction.

Main Results:

• FDCGAN significantly outperforms existing methods in restoring overexposed fringe images.
• The network successfully infers fringe structures and object shapes even when barely visible.
• Demonstrated enhanced quality of 3D reconstruction under extreme overexposure conditions.

Conclusions:

• FDCGAN offers a robust solution for fringe image restoration in challenging lighting conditions.
• The method has the potential to significantly improve measurement reliability in real-world FPP systems.
• Highlights the effectiveness of frequency-domain constraints and multi-scale learning for image restoration.