Two-dimensional phase unwrapping by a high-resolution deep learning network

Abstract

Phase measurements are of considerable interest in the field of optical metrology. Retrieving unambiguous phases from a single ambiguous wrapped phase map is a tricky problem. Conventional spatial phase unwrapping approaches are vulnerable to noise, the large variation range of phase surfaces, and phase discontinuity. In this study, the task of phase unwrapping is formulated as a pixel-wise semantic segmentation problem, and a high-resolution deep learning network named the HRNet is employed to infer an integer multiple of 2π for every pixel aiming to remove 2π discontinuities. The effectiveness and superiority of the proposed method are verified on both simulated and real data and through qualitative and quantitative comparisons with two deep learning-based methods and five classical spatial phase unwrapping methods.