Few-shot real-time quantitative phase imaging based on lightweight networks

Quantitative phase imaging plays a crucial role in fields such as environmental monitoring and atmospheric turbulence analysis. However, existing deep learning methods remain hampered by high computational complexity and heavy reliance on large-scale datasets, rendering them unsuitable for millisecond-scale real-time processing in demanding applications. Here, we propose a few-shot real-time quantitative phase imaging algorithm based on a lightweight network. The algorithm introduces what we believe to be a novel StarNet architecture that enables efficient feature extraction and dimensional expansion, significantly reducing computational overhead. Experimental results show that it can still provide phase reconstruction with millisecond-scale each frame, even under conditions of low signal-to-noise ratio and diverse random phase disturbances. This algorithm not only advances the practical deployment of real-time QPI technology but also offers an efficient and reliable solution for multi-frame inverse imaging problems in data-scarce scenarios.