物理模型引导的智能相位成像  

Intelligent phase imaging guided by physics models

作　　者：刘振 朱昊 周游 马展 曹汛[1] LIU Zhen;ZHU Hao;ZHOU You;MA Zhan;CAO Xun(School of Electronic Science and Engineering,Nanjing University,Nanjing 210023,China;Department of Computer Science and Technology,Nanjing University,Nanjing 210023,China;Medical School,Nanjing University,Nanjing 210093,China)

机构地区：[1]南京大学电子科学与工程学院,江苏南京210023 [2]南京大学计算机科学与技术系,江苏南京210023 [3]南京大学医学院,江苏南京210093

出　　处：《物联网学报》2023年第2期35-42,共8页Chinese Journal on Internet of Things

基　　金：国家自然科学基金资助项目(No.62025108,No.62071219,No.62101242)。

摘　　要：隐式神经表示使用神经网络刻画了信号坐标到其属性的映射,通过将正向物理过程引入误差函数的设计中,可被用于求解各种逆问题,应用前景广阔。然而,对隐式神经表示的网络参数进行随机初始化会导致优化速度慢、求解精度低,因此,提出使用元学习算法为隐式神经表示提供一个具有强先验的初始化参数,从而提升求解逆问题时的优化速度和求解精度。针对无透镜相位成像这一重要问题,基于快照式无透镜感知模型提出一种智能相位成像方法,将光学衍射传播理论引入隐式神经表示的误差函数设计中,能够消除传统深度学习算法对大规模数据集的依赖,仅需要传感器记录的单张强度图像,即可实现对样本的高精度相位恢复。此外,通过在网络初始化中引入元学习模型,进一步提升网络训练的效率和精度。数值仿真结果表明,与传统方法相比,所提方法能够获得11 dB以上的峰值信噪比(PSNR,peak signal-to-noise ratio)提升;在真实数据中的实验结果表明,所提方法重建出的相位图像更加清晰,伪影更少。Implicit neural representation characterizes the mapping between the signal’s coordinate to its attributes,and has been widely used in the optimization of inverse problems by embedding the physics process into the loss function.However,the implicit neural representation is suffering the low convergence speed and accuracy from the random initialization of the network parameters.The meta-learning algorithm for providing implicit neural representation with a strong prior of network parameters was proposed,thus enhancing the optimization efficiency and accuracy for solving inverse problems.To address the important issue of lens less phase imaging,an intelligent method on phase imaging was proposed based on the snapshot lens less sensing model.By embedding the optical diffraction propagation theory into the design of loss function for implicit neural representation,the dependency of large-scale labelled dataset in traditional deep learning-based methods was eliminated and accurate phase image from a single diffraction measurement image was provided.Furthermore,the meta-learning model was introduced for initializing network to further improve the efficiency and accuracy of network training.Numerical simulation results show that the proposed method can achieve a PSNR improvement of more than 11 dB compared to the conventional method.The experimental results in real data show that the phase image reconstructed by the proposed method is clearer and has fewer artifacts.

关 键 词：隐式神经表示 物理模型 相位成像 元学习 自监督学习 

分 类 号：TN26[电子电信—物理电子学]