Computational Imaging: Algorithms & Applications
Event details
| Date | 21.08.2019 |
| Hour | 09:00 › 11:00 |
| Speaker | Sepand Kashani-Akhavan |
| Location | |
| Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Pierre Vandergheynst
Thesis advisor: Prof. Martin Vetterli
Thesis co-advisor: Prof. Paul Hurley
Co-examiner: Prof. Jean-Philippe Thiran
Abstract
Signal and image analysis are key investigative techniques for extracting knowledge of natural
phenomena in various scientific fields. Often however the processes of interest cannot be directly
observed. Instead, indirect evidence of the phenomena responsible for image formation can be
measured and used to estimate the original image.
Advances in imaging science owe a large part to breakthroughs in instrumentation. Nevertheless the
traditional imaging landscape is challenged from all directions today: on the one hand applications
such as microscopy and astronomy constantly seek to resolve features beyond resolution limits of
their instruments. On the other hand medical imaging often has limited measurement time for
practical purposes and must learn to do more with less. Finally, conventional reconstruction
algorithms that work with current generation instruments may no longer be tractable at larger
scales. To bridge the gaps above, the trend is to involve computation in the imaging chain, by
combining knowledge of the acquisition system with mathematical models and optimization theory to
fuel the next revolution in imaging science.
This paper analyses three background works on classical and learning-based inverse problems in
computational imaging. We first look at recent work by members of the Event Horizon Telescope (EHT)
to image black holes through radio-interferometry using classical methods. Moving onto
learning-based methods, we look at design criteria to extend Convolutional Neural Networks (CNNs) to
non-Euclidean domains, specifically to discretized spherical manifolds. Finally, we revisit
classical methods with a focus on non-standard algorithmic solvers for compressed sensing problems.
Background papers
1. Super-resolution Interferometric Imaging with Sparse Modeling using
Total Squared Variation -- Application to Imaging the Black Hole Shadow
[Kuramochi, Akiyama, Ikeda, Tazaki, Fish, Pu, Asada, Honma].
2. DeepSphere: Efficient spherical Convolutional Neural Network with
HEALpix sampling for cosmological applications [Perraudin, Defferrard,
Kacprzak, Sgier].
3. Message-passing algorithms for compressed sensing [Donoho, Maleki,
Montanari].
Exam president: Prof. Pierre Vandergheynst
Thesis advisor: Prof. Martin Vetterli
Thesis co-advisor: Prof. Paul Hurley
Co-examiner: Prof. Jean-Philippe Thiran
Abstract
Signal and image analysis are key investigative techniques for extracting knowledge of natural
phenomena in various scientific fields. Often however the processes of interest cannot be directly
observed. Instead, indirect evidence of the phenomena responsible for image formation can be
measured and used to estimate the original image.
Advances in imaging science owe a large part to breakthroughs in instrumentation. Nevertheless the
traditional imaging landscape is challenged from all directions today: on the one hand applications
such as microscopy and astronomy constantly seek to resolve features beyond resolution limits of
their instruments. On the other hand medical imaging often has limited measurement time for
practical purposes and must learn to do more with less. Finally, conventional reconstruction
algorithms that work with current generation instruments may no longer be tractable at larger
scales. To bridge the gaps above, the trend is to involve computation in the imaging chain, by
combining knowledge of the acquisition system with mathematical models and optimization theory to
fuel the next revolution in imaging science.
This paper analyses three background works on classical and learning-based inverse problems in
computational imaging. We first look at recent work by members of the Event Horizon Telescope (EHT)
to image black holes through radio-interferometry using classical methods. Moving onto
learning-based methods, we look at design criteria to extend Convolutional Neural Networks (CNNs) to
non-Euclidean domains, specifically to discretized spherical manifolds. Finally, we revisit
classical methods with a focus on non-standard algorithmic solvers for compressed sensing problems.
Background papers
1. Super-resolution Interferometric Imaging with Sparse Modeling using
Total Squared Variation -- Application to Imaging the Black Hole Shadow
[Kuramochi, Akiyama, Ikeda, Tazaki, Fish, Pu, Asada, Honma].
2. DeepSphere: Efficient spherical Convolutional Neural Network with
HEALpix sampling for cosmological applications [Perraudin, Defferrard,
Kacprzak, Sgier].
3. Message-passing algorithms for compressed sensing [Donoho, Maleki,
Montanari].
Practical information
- General public
- Free