【作者】 陈龙； 杨义校； 陶然；

【Author】 CHEN Long;YANG Yixiao;TAO Ran;School of Information and Electronics,Beijing Institute of Technology;Beijing Key Laboratory of Fractional Signals and Systems;

【机构】 北京理工大学信息与电子学院； 分数域信号与系统北京市重点实验室；

【摘要】 相位恢复是一个具有挑战性的问题,它指从无相位测量中恢复所需信号,这广泛存在于计算成像领域。最近,由于深度学习方法能够隐式地学习图像先验知识,缓解相位恢复问题的病态性,基于深度学习的方法被应用于这一问题并取得了不错的成果。然而,一个关键问题仍未得到充分解答:训练好的神经网络在处理与训练时非常不同的相位恢复问题时,能泛化得多好?在实际应用中,这一点尤为重要,因为不同的成像设置、不同的物理测量和不同的应用都会导致不同的相位恢复问题。为了解决这个问题,一种受优化启发的神经网络架构被提出,该架构从用于相位恢复的近端梯度下降算法中衍生/展开,可以通过梯度下降利用物理模型信息,同时利用学习到的动态近端映射共同解决相位恢复问题。该方法的一个关键部分是开发一种动态近端映射模块,其参数可以在推理阶段动态调整,使其适应任何给定的成像设置。实验结果表明,该动态近端展开网络能够在多种成像条件下有效地从仅有幅度的测量中恢复丢失的相位,即使是在训练时未见过的情况下。更多还原

【Abstract】 Recovering the desired signal from phaseless measurements,phase retrieval(PR) is a challenging problem and widely encountered in the field of computational imaging.Recently,deep learning-based approaches have been applied to this problem with promising results,owing to its implicitly learned prior to alleviate the ill-posedness of the PR.However,a key question remains largely unanswered:how well can a trained neural network generalize to handle PR problems very different from the ones in training? This is particularly important in practice,since different imaging settings,different physical measurements,and different applications all lead to different PR problems.To address this concern,an optimization-inspired neural network architecture is proposed,unrolled from the proximal gradient descent algorithm for PR.It can exploit both the physical model information via gradient descent,as well as the image prior induced by the learned dynamic proximal mapping jointly to resolve the PR problem.A key part of the approach is to develop a dynamic proximal mapping module,whose parameters can be dynamically adjusted at the inference stage and make it adapt to any given imaging setting.Experimental results show that the dynamic proximal unrolling network can effectively retrieve missing phase from amplitude-only measurements under diverse imaging conditions,even those that are not seen during training.更多还原