EAGER: All-Optical Information Processing Device for Seeing Through Diffusers at the Speed of Light

项目编号：2054102首席研究员：Aydogan Ozcan （PI）和Mona Jarrahi （co-PI）机构：加州大学洛杉矶分校（University of California, Los angeles）： EAGER：通过漫射器以光速观察的全光信息处理设备项目描述：EAGER：电子、光子学和磁性设备非技术摘要：通过散射和漫射介质（如雾、云或人体组织）成像已经是几十年来的一个重要问题。毫无例外，之前所有的方法都是以数字计算机为核心的，因此信号首先由设备检测，然后用数字计算机处理以重建扩散器扭曲的图像。我们迫切需要新一代的光学设备，能够以光速看到、检测和量化目标物体，例如人体组织、墙壁、包裹、云、雾等，而不使用任何耗电的数字计算。这种独特的能力，一旦得到充分展示和开发，可能会在自主系统、生物医学成像、天文学、天体物理学、大气科学、安全、机器人和许多其他领域开辟各种新的应用。技术摘要：本文提出了一种不需要计算机的全光器件，它可以以光速看穿未知的漫射体，而不需要任何数字计算设备。与以前利用计算机重建未知扩散器后输入物体图像的数字方法不同，当扩散器扭曲的输入信号通过连续训练的衍射层衍射时，将使用一组衍射面/层创建一个被动设备，以全光学重建未知物体的图像，即图像重建将通过该设备以光速进行处理。设计的给定设备的每个衍射表面将具有数千个衍射特征（称为神经元），其中这些神经元的单个相位值将通过误差反向传播在训练阶段进行调整，通过最小化地面真实图像和输出视场中衍射图案之间的定制损失函数。在这种基于深度学习的衍射层设计之后，由此产生的无源器件将被制造成一个物理衍射光网络，该网络位于未知扩散器和输出/成像平面之间。当输入的物体光通过一个未知的漫射器时，散射光将被训练好的衍射装置收集，被动地重建畸变图像。该衍射器件的成功将在0.1-3太赫兹频段进行演示。与其他器件不同，所提出的衍射图像重建器件以光速运行，并且除了照明光外不需要任何功率。这种由无源衍射层实现的全光学图像重建将能够通过未知的扩散器看到物体，与现有的基于深度学习或迭代的计算机图像重建方法相比，它具有极低的功耗。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Proposal Number: 2054102Principal Investigator: Aydogan Ozcan (PI) and Mona Jarrahi (co-PI)Institution: University of California, Los AngelesTitle: EAGER: All-Optical Information Processing Device for Seeing Through Diffusers at the Speed of LightProgram Description: EAGER: Electronics, Photonics, and Magnetic DevicesNon-Technical Abstract:Imaging through scattering and diffusive media such as fog, clouds or human tissue has been an important problem for many decades. Without an exception, all the previous methods are based on, at their core, digital computers, such that the signals are first detected by a device and then processed using digital computers to reconstruct the diffuser-distorted images. There is an important and pressing need for a new generation of optical devices that can see, detect and quantify target objects through for example human tissues, walls, packages, clouds, fogs, etc., at the speed of light and without using any power-hungry digital computation. This unique capability, once fully demonstrated and developed, might open various new applications in autonomous systems, biomedical imaging, astronomy, astrophysics, atmospheric sciences, security, robotics, and many other fields.Technical Abstract:In this proposal, a computer-free, all-optical device that will see through unknown diffusers at the speed of light, without the need for any digital computation device will be developed. Unlike previous digital approaches that utilized computers to reconstruct an image of the input object behind unknown diffusers, a passive device will be created using a set of diffractive surfaces/layers to all-optically reconstruct the image of an unknown object as the diffuser-distorted input signals diffract through successive trained diffractive layers, i.e., the image reconstruction will be processed at the speed of light through this device. Each diffractive surface of a given device designed will have thousands of diffractive features (termed as neurons), where the individual phase values of these neurons will be adjusted in the training phase through error back-propagation, by minimizing a customized loss function between the ground truth image and the diffracted pattern at the output field-of-view. After this deep learning-based design of these diffractive layers, the resulting passive device will be fabricated to form a physical diffractive optical network that is positioned between an unknown diffuser and the output/image plane. As the input object light passes through an unknown diffuser, the scattered light will be collected by the trained diffractive device to passively reconstruct the distorted image. The success of this diffractive device will be demonstrated in 0.1-3 THz frequency band. Unlike other devices, the proposed diffractive image reconstruction device operates at the speed of light and does not require any power except for the illumination light. This all-optical image reconstruction that will be achieved by passive diffractive layers will enable to see objects through unknown diffusers and present an extremely low power device compared with existing deep learning-based or iterative image reconstruction methods implemented in computers.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.