Learning Pixel-wise Phase Unwrapping with Reinforcement Policy Gradient

Abstract This paper presents a novel method for pixel-wise phase unwrapping utilizing reinforcement learning (RL), addressing significant limitations of traditional techniques. Conventional phase unwrapping methods often struggle with noise and discontinuities, which can hinder their effectiveness in practical applications such as optical imaging and interferometry. In contrast, our approach employs Proximal Policy Optimization (PPO) to train an intelligent agent capable of making pixel-wise decisions based on the observed data, optimizing phase recovery without the need for extensive labeled datasets. We thoroughly evaluate our method on a variety of datasets, including both real and synthetic InSAR radar echo data, demonstrating its effectiveness across different scenarios. The experimental results reveal significant improvements in both accuracy and robustness compared to existing supervised learning techniques. Furthermore, our work not only advances the state-of-the-art in phase unwrapping but also underscores the potential of RL in solving complex imaging tasks. By paving the way for future applications in optical imaging and advanced interferometry, our research opens new avenues for exploring the capabilities of machine learning in handling intricate imaging challenges, thereby contributing to the broader field of computational imaging.