智能感知与计算研究中心
中国科学院自动化研究所   联系我们    English
 
    学术讲座

On the use of Artifical Intelligence for solving computational imaging problems

智能感知与计算系列讲座
Lecture Series in Intelligent Perception and Computing 

    TITLE): On the use of Artifical Intelligence for solving computational imaging problems

SPEAKER: George Barbastathis, Massachusetts Institute of Technology

(CHAIR)Dr. Yunlong Wang

    (TIME)Thursday (November. 15), 14:00 PM

    (VENUE) 1610 Meeting Room,16 Floor, Intelligent Building

报告摘要(ABSTRACT):

Computational Imaging systems consist of two parts: the physical part where light propagates through free space or optical elements such as lenses, prisms, etc. finally forming a raw intensity image on the digital camera; and the computational part, where algorithms try to restore the image quality or extract other type of information from the image data. A broad spectrum of computational imaging approaches exist: in one extreme, computer vision, the physical part typically comprises standard imaging optics; at the other extreme, computer vision, the physical part typically comprises standard imaging optics; at the other extreme, in lens-less imaging the burden of forming images or extracting other types of information from the optical field falls entirely on the computation.

In this talk I will discuss the emerging trend in computational imaging to train deep neural networks (DNNs) to perform image extraction and restoration tasks. In several imaging experiments carried out by our group, the objects rendered "invisible" due to various adverse conditions such as extreme defocus, scatter, or very low photon counts where "revealed" after processing of the raw images by DNNs. The DNNs were trained from examples consisting of pairs of known objects and their corresponding raw images. The objects were drawn from databases of faces and natural images, with the brightness converted to phase through a liquid-crystal spatial phase modulator. After training, the DNNs were capable of recovering unknown, i.e. hitherto not presented during training, objects from the raw images and recovery was robust to disturbances in the optical system, such as additional defocus or various misalignments. This suggests that DNNs may form robust internal models of the physics of light propagation and detection and generalize priors from the training set.

报告人简介(BIOGRAPHY):

George Barbastathis received the Diploma in Electrical and Computer Engineering in 1993 from the National Technical Computer Engineering in 1992 from the National Technical University of Athens and the MSc and PhD degrees in Electrical Engineering in 1994 and 1997, respectively, from the California Institute of Technology (Caltech.) After post-doctoral work at the University of Illinois at Urbana-Champaign, he joined the faculty at MIT in 1999, where he is now Professor of Mechanical Engineering. He has worked or held visiting appointments at Harvard University of Singapore, and the University of Michigan - Shanghai Jiao Tong University Joint Institute in Shanghai, People's Republic of China. His research interests are three-dimensional and spectral imaging; phase estimation; and gradient index optics theory and implementation with subwavelength-patterned dielectrics. He is member of the Institute of Electrical and Electronics Engineering (IEEE), and the American Society of Mechanical Engineers (ASME). In 2010 he was elected Fellow of the Optical Society of America (OSA) and in 2015 was a recipient of China's Top Foreign Scholar ("One Thousand Scholar") Award.


友情链接
 
中科院自动化研究所 智能感知与计算研究中心
中国科学院自动化研究所  事业单位   京ICP备14019135号-3