Pages

Thursday, November 20, 2008

CS: GPUCamp, a job, Super-Resolution using Finite Rate of Innovation.



Matthias Seeger has an opening for a PhD Student / Postdoctoral Researcher in Machine Learning, Image Processing

Probabilistic Machine Learning and Medical Image Processing, Saarland University, Saarbruecken, Germany

Fully-funded PhD/postdoc positions are available in the recently established Probabilistic Machine Learning group headed by Matthias Seeger (PhD). PhD training is conditional on acceptance to the International Max Planck Research School for Computer Science (based on evaluation of research proposal and oral presentation, after first six months).

Recent breakthroughs in large-scale approximate Bayesian inference for sparse continuous variable models allow nonlinear Bayesian experimental design (active learning) and compressed sensing to be applied to sampling optimization of magnetic resonance imaging. Details about these projects.

Saarland University is among the leading computer science faculties in Europe, with world-class groups in computer graphics, theory of algorithms and programming languages, theoretical CS, and bioinformatics, among others. It features a unique accumulation of top-ranked CS research institutes (Max Planck Institute for Informatik, Max Planck Institute for Software Systems, DFKI). Within the recently established interdisciplinary MMCI Cluster of Excellence, 20 independent research groups are working in areas with strong overlaps to core machine learning application areas. Saarbruecken is dedicated to excellent postgraduate education, structured according to international standards in the International Max Planck Research School for Computer Science (courses taught in english).

The Probabilistic Machine Learning group focusses on theory and applications of approximate Bayesian inference, and its scalable reduction to standard methods of scientific computing (numerical mathematics, efficient algorithms, signal processing, parallel processing). We closely collaborate with the Center for High-field Magnetic Resonance, Max Planck Institute for Biological Cybernetics, Tuebingen (with a range of MR scanners dedicated to basic research), and have close ties to the Empirical Inference group (headed by Bernhard Schoelkopf) at the same institute, beyond close connections to top machine learning groups in the UK and US.

We are looking for highly motivated, research-oriented individuals with an excellent grasp of the mathematics underlying approximate Bayesian inference, or/and numerical optimization and mathematics, or/and image and signal processing. A strong theoretical background in a field relevant to analysis of statistical methods, or/and keen interest and capabilities in large-scale scientific programming are required.

Please be sure to include the following in your application:

  • Curriculum vitae
  • Statement of research interests (1 page)
  • Letters of reference (1-3) from referees able to comment on your work and academic standing (PhD/MSc thesis advisor, supervisor for internships)
  • Sample of your strongest work (first-author paper in peer-reviewed journal/conference, MSc or PhD thesis, term project paper (with official record attesting your authorship)) in the rough area of interest
  • Transcript of studies (for PhD applicants)

Applications should be sent by e-mail to Matthias Seeger. If you happen to attend the forthcoming Neural Information Processing Systems conference, please make yourself known to me there.

Matthias Seeger goes further in the description of his projects, please note the question in the compressive sensing section:

Currently, a number of PhD/postdoc positions are available for exceptional candidates with strong interests in applications of approximate Bayesian inference. Aspects on several levels of these projects lead into previously rather unchartered terrain, with much original work still to be done:

  • Supporting MRI sequence design by Machine Learning
    Some ML work has been done on analyzing MR images (mostly fMRI) after they have been recorded, or on denoising images. Our interest is in supporting decisions about how images are measured in the first place. On the other hand, we will also identify and address applications of Machine Learning and Bayesian techniques to MRI problems other than sequence design (for example, estimation and compensation of gradient or main field inhomogeneities, robust phase difference estimation, combination of measurements from coil arrays, or motion correction).
  • Nonlinear (Bayesian) Experimental Design
    Bayesian (or classical) ED for Gaussian MRFs is well-developed, but we are not aware of work on the scale of interest here for non-Gaussian models, which are much more useful in practice for a number of reasons. Moreover, there seem to be no theoretical results about Bayesian ED with variational posterior approximations.
  • Compressed sensing for images (that really works)
    Compressed sensing is very fashionable right now in signal processing, and MRI is often cited as a useful application. However, as noted above, present theory has little relevance for the practical problem. What are the relevant properties of a good (or optimal) design in the context of sparse reconstruction of real MR images?
  • Large scale approximate inference for sparse generalized linear models
    Bayesian ED needs inference, beyond penalized optimization for the posterior mode. This seems more difficult, and is almost untouched by theoretical statistics so far (while more and more results about sparse penalized estimation are obtained). Our relaxation is provably convex, inviting theoretical analysis in principle.
    Moreover, our novel algorithms, orders of magnitude faster than previous methods for inference over images, can be applied to generalized linear models as well, opening up a host of new applications of Bayesian inference beyond MAP estimation.

The annoucement will be added to the Compressive Sensing Job section.

and also I found two papers using Finite Rate of Innovation:

Exact Feature Extraction using Finite Rate of Innovation Principles with an Application to Image Super-resolution by Loic Baboulaz and Pier Luigi Dragotti. The abstract reads:

The accurate registration of multiview images is of central importance in many advanced image processing applications. Image super-resolution, for example, is a typical application where the quality of the super-resolved image is degrading as registration errors increase. Popular registration methods are often based on features extracted from the acquired images. The accuracy of the registration is in this case directly related to the number of extracted features and to the precision at which the features are located: images are best registered when many features are found with a good precision. However, in low-resolution images, only a few features can be extracted and often with a poor precision. By taking a sampling perspective, we propose in this paper new methods for extracting features in low resolution images in order to develop efficient registration techniques. We consider in particular the sampling theory of signals with infinite rate of innovation [10] and show that some features of interest for registration can be retrieved perfectly in this framework, thus allowing an exact registration. We also demonstrate through simulations that the sampling model which enables the use of infinite rate of innovation principles is well-suited for modeling the acquisition of images by a camera. Simulations of image registration and image super-resolution of artificially sampled images are first presented, analyzed and compared to traditional techniques. We finally present favorable experimental results of super-resolution of real images acquired by a digital camera available on the market.

Some videos can be found on Loic Baboulaz's page.



Looking back at the images we took from a NASA stratospheric balloons, it looks like I do not have a high concentration of images for this implementation (40). All the panoramas are listed here, original photos from the flight can be found here. The camera worked for several hours but we were limited by the 4GB SD card. More information can be found here and here. I launched this student project as a direct response to the deplorable rescue effort that occured after Katrina. All these photos are freely available for all your benchmarks with attribution.

Geometry-Driven Distributed Compression of the Plenoptic Function: Performance Bounds and Constructive Algorithms by Nicolas Gehrig and Pier Luigi Dragotti. The abstract reads:

In this paper, we study the sampling and the distributed compression of the data acquired by a camera sensor network. The effective design of these sampling and compression schemes requires, however, the understanding of the structure of the acquired data. To this end, we show that the a-priori knowledge of the configuration of the camera sensor network can lead to an effective estimation of such structure and to the design of effective distributed compression algorithms. For idealized scenarios, we derive the fundamental performance bounds of a camera sensor network and clarify the connection between sampling and distributed compression. We then present a distributed compression algorithm that takes advantage of the structure of the data and that outperforms independent compression algorithms on real multi-view images.


Credit Photo: Franky De La Garza, Jay Gerber, Sara Guest, Raymond Mendoza, Ramon Rivera, Karen Villatoro, Pamela Withrow, John Yezak, Igor Carron, Pedro Davalos.

No comments:

Post a Comment