基于深度学习的动态三维数字全息像重建方法研究

Digital holography is a significant imaging technique that can record the whole wavefront information of a three-dimensional (3D) object in real time. It is one of the hot topics in the optical imaging field since it can provide a simple but effective way for microscopic imaging and quantitative measurement. For the quantitative phase imaging of a dynamic 3D target, however, conventional holographic reconstruction methods can only handle one single 2D hologram in a sequential way. Afterwards, the stack of 2D reconstructed images is synthesized to a 3D volume. Although functional, the process is computational demanding and the quality is barely satisfactory. In addition, the underlying relationships between frames (i.e., holograms) are not taken into consideration and not utilized sufficiently. Both of the cumbersome operations and neglect of hidden information lead to unprecedented difficulties in the development of real-time applications with digital holography. In recent years, deep learning, or artificial intelligence, is an emerging powerful tool for image processing. It provides a promising solution since it can not only deal with 2D images, but also process video streams in an end-to-end manner. This project aims at studying the enhancement and optimization of reconstruction methods by making use of deep learning in digital holographic imaging. The three main targets that will be involved in are: 1. Design a generalized deep learning-based reconstruction method and achieve a generalized physical realization for quantitative phase imaging; 2. Design an end-to-end 3D reconstruction model for holographic 3D image reconstruction; 3. Based on high-dimensional deep learning, achieve real-time object recognition, holographic image reconstruction and tracking in dynamic 3D multi-sample scenarios. The research achievements will reveal the mechanism of applying deep learning to optical imaging. This project will also benefit to novel real-time holographic image reconstruction methods and pave the way to biological imaging and object detection in digital holography.

数字全息术能够实时记录三维物体的全部波前信息，为显微成像、定量检测提供了一种简单而有效的实现途径，因此成为光学成像领域的研究热点之一。然而对于动态三维目标的定量相位成像，传统的全息像重建算法只能对每一帧二维全息图分别重建、再进行三维合成。这种方法不但耗时长、效果差，也未充分利用帧与帧之间的内在联系，特别不利于实时动态全息应用场景的开发。深度学习不仅可以处理二维图像，还能对视频信息实现端到端式处理，有助于解决上述问题。本项目旨在研究深度学习对数字全息术中全息像重建算法的增强与优化。主要研究内容：①通用型深度学习网络架构建立及通用型定量相位成像的物理实现; ②三维物体实现端到端式三维全息像的重建方法; ③动态多目标实时检测、全息像重建以及追迹的高维度深度学习网络架构。研究成果将有助于揭示深度学习在光学成像应用中的新机制，并为发展新型实时三维全息像重建算法及其在生物目标检测中的应用提供新的途径。

数字全息术能够实时记录三维物体的全部波前信息，为显微成像、定量检测提供了一种简单而有效的实现途径，因此成为光学成像领域的研究热点之一。然而对于动态三维目标的定量相位成像，传统的全息像重建算法只能对每一帧二维全息图分别重建、再进行三维合成。这种方法不但耗时长、效果差，也未充分利用帧与帧之间的内在联系，特别不利于实时动态全息应用场景的开发。深度学习不仅可以处理二维图像，还能对视频信息实现端到端式处理，有助于解决上述问题。本项目旨在研究深度学习对数字全息术中全息像重建算法的增强与优化。主要研究内容：①通用型深度学习网络架构建立及通用型定量相位成像的物理实现; ②三维物体实现端到端式三维全息像的重建方法; ③动态多目标实时检测、全息像重建以及追迹的高维度深度学习网络架构。在本项目的支持下，共发表期刊论文12篇，其中包括APL Photonics一篇，Optics Express三篇，Optics and Lasers in Engineering三篇，IEEE Photonics Journal一篇，IEEE Access一篇，《光学学报》两篇，《激光与光电子学进展》一篇；会议论文2篇；已提交专利申请2项。相关研究成果将有助于揭示深度学习在光学成像应用中的新机制，并为发展新型实时三维全息像重建算法及其在生物目标检测中的应用提供新的途径。