2D phase unwrapping by attention-guided deep neural network for optical interferometry

Two-dimensional phase unwrapping is an essential process in interferometry applications since the unwrapped phase plays a guiding role in the morphological reconstruction of the object. However, the unwrapped phase at each position is affected by the wrap states for other positions. Therefore, it is essential to obtain global context information for better phase unwrapping. To this purpose, we propose a novel attention-guided deep neural network that introduces the self-attention mechanism, a popular deep learning technique, to tackle the problem in a highly efficient way. We design a dual attention phase unwrapping network (DAPUN), which uses two kinds of complementary self-attention structures to obtain global context information (such as wrapped states), allowing each positional feature to be directly compared with ones at any other positions. To evaluate the effectiveness of the proposed DAPUN, we simulate extensive data to train the networks and compare it with several existing phase unwrapping methods. The results show that DAPUN significantly outperforms the previous state-of-the-art. After that, we apply the trained DAPUN to the real cases to demonstrate its generalization capability.