2D Phase Unwrapping via Graph Cuts

Abstract

Phase imaging technologies such as interferometric synthetic aperture radar (InSAR), magnetic resonance imaging (MRI), or optical interferometry, are nowadays widespread and with an increasing usage. The so-called phase unwrapping, which consists in the in- ference of the absolute phase from the modulo-2π phase, is a critical step in many of their processing chains, yet still one of its most challenging problems. We introduce an en- ergy minimization based approach to 2D phase unwrapping. In this approach we address the problem by adopting a Bayesian point of view and a Markov random field (MRF) to model the phase. The maximum a posteriori estimation of the absolute phase gives rise to an integer optimization problem, for which we introduce a family of efficient algo- rithms based on existing graph cuts techniques. We term our approach and algorithms PUMA, for Phase Unwrapping MAx flow. As long as the prior potential of the MRF is convex, PUMA guarantees an exact global solution. In particular it solves exactly all the minimum L p norm (p ≥ 1) phase unwrapping problems, unifying in that sense, a set of existing independent algorithms. For non convex potentials we introduce a version of PUMA that, while yielding only approximate solutions, gives very useful phase unwrap- ping results. The main characteristic of the introduced solutions is the ability to blindly preserve discontinuities. Extending the previous versions of PUMA, we tackle denoising by exploiting a multi-precision idea, which allows us to use the same rationale both for phase unwrapping and denoising. Finally, the last presented version of PUMA uses a frequency diversity concept to unwrap phase images having large phase rates. A representative set of experiences illustrates the performance of PUMA.