Phase Unwrapping via Factor Graph Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for unwrapping noisy two-dimensional phase-wrapped images, such as those in SAR and MRI, often fail to effectively remove noise and constrain phase derivatives to satisfy the zero-curl constraint, leading to inefficient and inaccurate phase recovery.
Innovation Solution
A method using a factor graph with variable nodes representing integer-shifts and re-wrapped phase values, along with factor nodes for noise, smoothness, and zero-curl constraints, iteratively optimizes these values through minimum cost flow and iterated conditional modes procedures to produce the most probable absolute phase image.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional phase-unwrapping methods are used, then the unwrapping process can be completed, but noise removal is ineffective and the zero-curl constraint is not satisfied
Solution Approach 1:
The phase-unwrapping problem is segmented into two distinct optimization tasks: integer-shift determination and continuous-phase optimization. This is achieved by introducing auxiliary variables that separate the discrete unwrapping operation from the continuous phase adjustment, allowing each to be optimized independently through specialized algorithms (minimum-cost flow for integer shifts, gradient descent for continuous phases).
Solution Approach 2:
Auxiliary variables are introduced as intermediary elements between the wrapped phase measurements and the final unwrapped phase solution. These auxiliary variables represent the integer shifts and continuous phase adjustments, mediating the transformation process and enabling the satisfaction of both noise removal and zero-curl constraints through structured optimization.
2Productivity
If traditional phase-unwrapping methods are used, then processing can be completed, but the computational efficiency is insufficient for large images
Solution Approach 1:
The algorithm is segmented into distinct computational stages: integer-shift optimization using minimum-cost flow, followed by continuous-phase optimization using gradient descent. This segmentation allows each stage to use algorithms optimized for its specific mathematical structure, improving overall computational efficiency for large images despite the increased algorithmic complexity.
Solution Approach 2:
The approach changes the parameterization of the phase-unwrapping problem by introducing auxiliary variables for integer shifts and continuous phases. This parameter transformation enables the use of efficient optimization algorithms tailored to each parameter type, significantly improving processing speed for large-scale images compared to traditional monolithic approaches.
3Ease of manufacture
If the zero-curl constraint is not enforced, then the unwrapping process is simpler, but the phase derivatives do not satisfy physical consistency
Solution Approach 1:
The zero-curl constraint is enforced through auxiliary variables that act as intermediaries between the phase measurements and the unwrapped solution. These variables are optimized subject to explicit zero-curl constraints, ensuring physical consistency while maintaining algorithmic tractability through the structured optimization framework.
Solution Approach 2:
The zero-curl constraint is built into the optimization framework from the beginning, rather than being applied as a post-processing step. The auxiliary variables are optimized subject to zero-curl constraints throughout the iterative process, ensuring physical consistency is maintained throughout the unwrapping procedure.
Data Source
AI summary
A method converts an input image of noisy wrapped phases to an output image of absolute unwrapped phases. The noisy wrapped phases in the input image are represented as a set of re-wrapped phases and a set of phase shifts. The set of re-wrapped phases are partitioned into a first group and a second group. Integer differences between the phase shifts are optimized while holding the re-wrapped phases fixed. Then, the first group of re-wrapped phases are optimized, while holding the integer differences between the phase shifts, and the second group of re-wrapped phases fixed. The integer differences between the phase shifts are optimized again, while holding the re-wrapped phases fixed. Then, the second group of re-wrapped phases are optimized, while holding the integer differences between the phase shifts, and the first group of re-wrapped phases fixed. The optimizing steps are repeated until the re-wrapped phase converge. Then, the converged re-wrapped phases and integer differences between the phase shifts are output as an output image of absolute unwrapped phases.