Two-dimensional phase unwrapping and denoising network based on frequency-domain operators and sparse attention

Area of Science:

• Image processing
• Computer vision
• Signal processing

Background:

• Two-dimensional phase unwrapping is essential for fields like optical interferometry and synthetic aperture radar.
• Traditional phase unwrapping methods struggle with noise and phase discontinuities, leading to errors.

Purpose of the Study:

• To develop a robust and accurate deep learning model for two-dimensional phase unwrapping.
• To address the limitations of conventional methods in handling complex data conditions.

Main Methods:

• Proposed FSA-PU, a lightweight U-Net-based deep learning model.
• Integrated multi-scale context modeling, sparse global dependency, and frequency-domain reconstruction.
• Utilized depthwise separable convolutions for improved edge and detail recovery.

Main Results:

• FSA-PU demonstrated superior performance on synthetic and real-world datasets.
• Achieved excellent metrics including Peak Signal-to-Noise Ratio (PSNR), Root Mean Square Error (RMSE), and Structural Similarity Index Measure (SSIM).
• Showcased high accuracy and robustness, particularly under challenging noise and discontinuity conditions.

Conclusions:

• FSA-PU offers a significant advancement in two-dimensional phase unwrapping.
• The model provides a reliable solution for applications requiring precise phase information from noisy or discontinuous data.