Physics based modeling for the development of soft segmentation and reconstruction algorithms

We propose a new regularization approach, using a soft-spin, phase-field model prior, that incorporates physics-based parameter estimates for joint soft segmentation and for reconstruction algorithms that avoid discontinuities between region labels. The phase-field model employed here is based on a continuous, coarse-grained field variable describing distinct crystallites having different orientations that comprise a poly-crystalline microstructure. The proposed algorithm, called Snap is used here to perform joint segmentation and reconstruction of materials microstructures from several noisy numerical measurements. Comparisons with state-of-the-art de-noising algorithms demonstrate the superior performance of Snap.