Using neural networks to create a reliable phase quality map for phase unwrapping

Area of Science:

• Interferometric Signal Processing
• Computational Imaging
• Machine Learning Applications

Background:

• Two-dimensional phase unwrapping is essential for interferometric signal processing.
• Existing quality maps struggle with noisy, low-quality data and direct gradient quality assessment.
• Current methods often approximate gradient quality, impacting network-based unwrapping algorithms.

Purpose of the Study:

• To develop a direct quality map generation method for two-dimensional phase unwrapping.
• To improve the handling of low-quality and fast-changing regions in interferometric data.
• To enhance the performance of both path-based and network-based phase unwrapping algorithms.

Main Methods:

• Analysis of the fundamental properties of quality maps in phase unwrapping.
• Proposal of a convolutional neural network (CNN) based approach for quality map generation.
• Generation of a pair of quality maps, specifically for horizontal and vertical gradients.

Main Results:

• The proposed CNN method generates quality maps that directly represent gradient quality.
• The generated quality maps effectively address limitations of existing methods in noisy regions.
• Experimental validation demonstrates improved performance of phase unwrapping algorithms using the new quality maps.

Conclusions:

• The novel CNN-based quality map generation method offers a significant advancement in two-dimensional phase unwrapping.
• This approach enhances the robustness and accuracy of phase unwrapping, particularly in challenging data conditions.
• The generated quality maps provide a more direct and effective means to guide phase unwrapping algorithms.