Deep learning techniques are always bound with big data and large, sophisticated models. In this paper, we show that this is not necessarily true for the task of end-to-end phase retrieval in off-axis interferometric quantitative phase imaging. For this task, we first introduce a new loss function, called bucket error rate (BER), for addressing the problem of imbalanced data distribution by balancing loss-bias of target and background area adaptively. With BER, we demonstrate that a U-Net model can learn the underneath logic for converting a raw interferogram to a phase map from only one training sample. At last, we present a novel mixed-context network (MCN) which can simultaneously aggregate local- and global-contextual information. Experimental results show that compared to U-Net, the proposed MCN is more accurate, more compact, and can be trained faster.
|