IEEE 2017 NSS/MIC/RTSD ControlCenter

Online Program Overview Session: BDW-01

To search for a specific ID please enter the hash sign followed by the ID number (e.g. #123).

Big Data Workshop

Session chair: Marcel Demarteau Argonne National Laboratory
Shortcut: BDW-01
Date: Friday, October 27, 2017, 08:00
Room: Hannover A&B
Session type: Workshop


8:00 am BDW-01-1

3D Computational Phase Microscopy (#1202)

M. Chen1, D. Parkinson2, L. Waller1

1 UC Berkeley, Electrical Engineering and Computer Sciences, Berkeley, California, United States of America
2 Lawrence Berkeley National Labs, Berkeley, California, United States of America


Computational imaging involves the joint design of imaging system hardware and software, optimizing across the entire pipeline from acquisition to reconstruction. This talk will describe new methods for computational phase microscopy of thick 3D samples in both the optical and X-ray regimes. We use computational illumination and detection, combined with simple and inexpensive hardware modifications and advanced image reconstruction algorithms to achieve fast, robust imaging of 3D refractive index maps. Our algorithms are based on large-scale nonlinear non-convex optimization procedures for phase retrieval, with appropriate priors.

Keywords: phase retrieval, computational imaging, 3D imaging, image reconstruction
8:36 am BDW-01-2

X-ray Fourier-transform Ghost Imaging via Sparsity Constraints (#2412)

H. Yu1, R. Lu1, Z. Tan1, R. Zhu1, S. Han1

1 Shanghai Institute of Optics and Fine Mechanics, Chinese Academy of Sciences, Jiading, Shanghai, China


Conventional imaging techniques acquire information of an object based on the point-to-point correspondence between the object-space and the image-space. It usually results in data redundancy, and the information acquisition efficiency is much lower than the Shannon Limit determined by Information Theory. However, in ghost imaging, the information of an object is coded by the fluctuations of light field. Combining with compressive sensing theory, ghost imaging via sparsity constraints (GISC) provides the way of approaching the Shannon Limit, and it has many potential applications, such as three-dimensional lidar and spectral camera in real space, x-ray Fourier-transform diffraction imaging in reciprocal space, etc.

The experimental demonstration of x-ray Fourier-transform ghost imaging via sparsity constraints has been fulfilled recently, which may extend x-ray crystallography to non-crystalline samples. However, there are still some obstacles in the road to x-ray GISC nanoscope. Here we present a tabletop microscopy system based on the x-ray Fourier-transform GISC method. Since efficient information coding and sampling is essential in GISC, we optimize the fluctuations of x-ray field to achieve a high quality pseudothermal source and propose a data-multiplexing algorithm. Researches show that x-ray speckle patterns with better contrast in a large area can be obtained by positioning an appropriate pinhole array in front of a designed diffuser, and the sampling efficiency can be greatly improved by the multiplexing of the intensities at different detecting positions.

Further work in progress concentrates on the enhancement of imaging sensitivity, the improvement of phase recovery algorithms, and taking full advantage of the prior knowledge to obtain more high-frequency information of nanoscale samples.

Keywords: x-ray imaging, ghost imaging, compressive sensing
9:12 am BDW-01-3

Reducing the data acquisition time of synchrotron X-ray imaging with a deep learning approach (#1738)

X. Yang1, V. De Andrade1, D. Gürsoy1, F. De Carlo1

1 Argonne National Laboratory, X-ray Science Division, Lemont, Illinois, United States of America


During the last decade, the use of new generation synchrotron facilities with ultra-high brilliance X-ray sources combined with the development of fast detectors have enabled dynamic X-ray imaging. However, some systems are changing so quickly that characterization of their time evolution is still out of reach. The X-ray images with too short exposure time show strong noise and less structural information of the objects. It is expensive to improve the quality of the dynamic X-ray images by improving the measurement setups, and sometimes is far beyond the capability for current hardware. Computational methods are cheap and easy options. Traditional image process is powerless for most of the X-ray imaging results, because the complex of the image features and image qualities. The deep convolutional neural network (CNN) is a feature based image analyzing method. It shows great potential to model the complex image problems such as X-ray imaging. We will present our recent progress of applying the CNN on synchrotron X-ray imaging to reduce the data acquisition time by enhancing the X-ray images of short exposure time. We developed a CNN algorithm to learn the transformation rule between X-ray images with short exposure time and long exposure time. Our algorithm has been quantitatively evaluated with simulation data. We also validated it with nano-CT measurements of a mouse brain sample and battery particle sample. The results show that our CNN method is able to speed up X-ray image acquisitions by at least a factor of 10 without losing the structural information and generating artifacts.

Keywords: X-ray imaging, reducing data acquisition time, deep learning, convolutional neural network