Deep neural network for fringe pattern filtering and normalization.

Area of Science:

• Optics and Photonics
• Computer Vision
• Machine Learning

Background:

• Fringe patterns (FPs) are crucial in optical metrology but are susceptible to noise and require normalization.
• Existing methods for FP processing often struggle with complex noise distributions and require manual parameter tuning.

Purpose of the Study:

• To introduce a novel deep learning framework for automated denoising and normalization of fringe patterns.
• To adapt the U-net architecture for fringe pattern processing tasks, enhancing reconstruction quality.

Main Methods:

• Utilized the U-net neural network architecture for fringe pattern normalization.
• Developed a modified U-net, termed V-net, with adjusted weight distribution for improved reconstruction.
• Proposed residual (ResV-net) and fast operating versions of V-net to explore performance enhancements.

Main Results:

• The V-net model demonstrated high-quality results in fringe pattern filtering and normalization.
• Evaluated performance across various noise levels and distributions on synthetic and real fringe pattern data.
• Experimental results confirmed the V-net's superiority over state-of-the-art methods in fringe pattern processing.

Conclusions:

• The proposed deep learning framework, particularly the V-net, offers a powerful and effective paradigm for fringe pattern processing.
• This approach shows significant potential for improving the accuracy and efficiency of interferogram analysis.
• The study highlights the capabilities of deep neural networks in learning complex FP characteristics for denoising and normalization.