Spatial Relation Awareness Module for Phase Unwrapping

Abstract

Phase unwrapping is a technique used to recover the original phase from the wrapped phase in the range (−π,π]. Various methods have been proposed for phase unwrapping. In particular, methods using convolutional neural networks (CNNs) have been extensively researched because of their high robustness against noise and fast inference speed. However, conventional CNN-based methods discard the local position information and relationships between pixels in the convolution process, resulting in poor phase-unwrapping performance. To obtain better phase unwrapping results, we propose a module that combines a global convolution network, which applies convolutional layers with a kernel size equivalent to that of the feature maps, and CoordConv, which acquires the positional relationships between pixels. We validated the performance of the proposed method by comparing it with a quality-guided path algorithm and deep learning-based phase unwrapping methods and found that the proposed method is highly robust against noise.