Two-dimensional phase unwrapping based on U²-Net in complex noise environment.

Area of Science:

• Computer Vision and Image Processing
• Deep Learning for Scientific Applications

Background:

• Conventional two-dimensional phase unwrapping (PU) methods are constrained by the Itoh condition.
• Deep learning techniques show promise in overcoming PU limitations but often lack testing in complex, real-world noise environments.
• The performance differences of various deep learning models under diverse noise types and phase changes remain underexplored.

Purpose of the Study:

• To evaluate and compare the performance of several deep learning models for two-dimensional phase unwrapping (PU).
• To assess the robustness of these models against additive Gaussian white noise and multiplicative speckle noise.
• To identify the most effective deep learning architecture for PU in complex and noisy real-world scenarios.

Main Methods:

• Application of the nested U²-Net architecture for two-dimensional phase unwrapping.
• Comparative analysis of U-Net, DLPU-Net, VUR-Net, PU-GAN, U²-Net, and U²-Netp.
• Simulation of complex noise environments using Gaussian white and speckle noise; testing on 421 real-phase images (candle flames, pits, grooves, tables).

Main Results:

• The U²-Net, featuring U-like residual blocks, demonstrated superior anti-noise performance and structural stability.
• U²-Net and the lightweight U²-Netp exhibited higher accuracy, enhanced noise resistance, and better generalization abilities.
• Qualitative evaluation considered model parameters, floating-point operations, and PU speed; quantitative evaluation used MSE, PSNR, and SSIM.

Conclusions:

• The U²-Net architecture offers significant advantages for two-dimensional phase unwrapping, particularly in noisy conditions.
• U²-Net and U²-Netp provide a robust and accurate solution for phase unwrapping, outperforming existing deep learning models.
• The findings highlight the potential of U²-Net for practical applications requiring reliable phase unwrapping from complex, real-world data.