EPUN: Efficient 2D Phase Unwrapping with a Lightweight Neural Network

Phase unwrapping is a fundamental task. In recent years, deep learning has been widely used for phase unwrapping in different fields. Nevertheless, along with the unwrapping performance improvement of deep networks, their structures become more and more complicated and diverse. As a result, these designed networks make it difficult to achieve an optimal balance between efficiency and effectiveness. For this purpose, we proposed a lightweight neural network for efficient 2d phase unwrapping (EPUN). EPUN first introduces a feature modulation block to efficiently aggregate long-range spatial context while executing channel mixing on the features. In addition, we design a feature supervision module exclusively used during training, which provides explicit unwrapping supervision for deep-level features, facilitating the network to converge to a better parameter set for optimized phase unwrapping performance. We evaluate EPUN qualitatively and quantitatively on both simulation and real data. The experimental results show that EPUN not only achieves optimal unwrapping performance but also has excellent unwrapping speed.