光学学报, 2020, 40 (1): 0111003, 网络出版: 2020-01-06   

深度学习下的计算成像:现状、挑战与未来 下载: 11204次特邀综述

Deep Learning Based Computational Imaging: Status, Challenges, and Future
左超 1,2冯世杰 1,2张翔宇 1,2韩静 2陈钱 2,*
作者单位
1 南京理工大学电子工程与光电技术学院,智能计算成像实验室(SCILab), 江苏 南京 210094
2 南京理工大学江苏省光谱成像与智能感知重点实验室, 江苏 南京 210094
摘要
近年来,光学成像技术已经由传统的强度、彩色成像发展进入计算光学成像时代。计算光学成像基于几何光学、波动光学等理论对场景目标经光学系统成像再到探测器采样这一完整图像生成过程建立精确的正向数学模型,再求解该正向成像模型所对应的“逆问题”,以计算重构的方式来获得场景目标的高质量图像或者传统技术无法直接获得的相位、光谱、偏振、光场、相干度、折射率、三维形貌等高维度物理信息。然而,计算成像系统的实际成像性能也同样极大程度地受限于“正向数学模型的准确性”以及“逆向重构算法的可靠性”,实际成像物理过程的不可预见性与高维病态逆问题求解的复杂性已成为这一领域进一步发展的瓶颈问题。近年来,人工智能与深度学习技术的飞跃式发展为计算光学成像技术开启了一扇全新的大门。不同于传统计算成像方法所依赖的物理驱动,深度学习下的计算成像是一类由数据驱动的方法,它不但解决了许多过去计算成像领域难以解决的难题,还在信息获取能力、成像的功能、核心性能指标(如成像空间分辨率、时间分辨率、灵敏度等)上都获得了显著提升。基于此,首先概括性介绍深度学习技术在计算光学成像领域的研究进展与最新成果,然后分析了当前深度学习技术在计算光学成像领域面临的主要问题与挑战,最后展望了该领域未来的发展方向与可能的研究方向。
Abstract
In recent years, optical imaging techniques have entered into the era of computational optical imaging from the traditional intensity and color imaging. Computational optical imaging, which is based on geometric optics, wave optics, and other theoretical foundations, establishes an accurate forward mathematical model for the whole image formation process of the scene imaged through the optical system and then sampled by the digital detector. Then, the high-quality reconstruction of the image and other high dimensional information, such as phase, spectrum, polarization, light field, coherence, refractive index, and three-dimension profile, which cannot be directly accessed using traditional methods, can be obtained through computational reconstruction method. However, the actual imaging performance of the computational imaging system is also limited by the “accuracy of the forward mathematical model” and “the reliability of inverse reconstruction algorithm”. Besides, the unpredictability of real physical imaging process and the complexity of solving high dimensional ill-posed inverse problems have become the bottleneck of further development of this field. In recent years, the rapid development of artificial intelligence and deep learning for the technology opens a new door for computational optical imaging technology. Unlike “physical driven” model that traditional computational imaging method is based on, computational imaging based on deep learning is a kind of “data-driven” method, which not only solves many problems considered quite challenge to be solved in this field, but also achieves remarkable improvement in information acquisition ability, imaging functions, and key performance indexes of imaging system, such as spatial resolution, temporal resolution, and detection sensitivity. This review first briefly introduces the current status and the latest progress of deep learning technology in the field of computational optical imaging. Then, the main problems and challenges faced by the current deep learning method in computational optical imaging field are discussed. Finally, the future developments and possible research directions of this field are prospected.

左超, 冯世杰, 张翔宇, 韩静, 陈钱. 深度学习下的计算成像:现状、挑战与未来[J]. 光学学报, 2020, 40(1): 0111003. Chao Zuo, Shijie Feng, Xiangyu Zhang, Jing Han, Chen Qian. Deep Learning Based Computational Imaging: Status, Challenges, and Future[J]. Acta Optica Sinica, 2020, 40(1): 0111003.

本文已被 35 篇论文引用
被引统计数据来源于中国光学期刊网
引用该论文: TXT   |   EndNote

相关论文

加载中...

关于本站 Cookie 的使用提示

中国光学期刊网使用基于 cookie 的技术来更好地为您提供各项服务,点击此处了解我们的隐私策略。 如您需继续使用本网站,请您授权我们使用本地 cookie 来保存部分信息。
全站搜索
您最值得信赖的光电行业旗舰网络服务平台!