Two applications of sparse and low-rank signal modelling

Thumbnail

Event details

Date 27.03.2014
Hour 13:3014:30
Speaker Dr. Pavel Rajmic, SPLAB, Brno University of Technology, Czech Republic
Bio: Pavel Rajmic has been employed as researcher at the Faculty of Electrical Engineering and Communication, Brno University of Technology since 2004. He is a member of the team SPLab dealing with digital multimedia signals processing at the Department of Telecommunications. In 2001, he worked as a statistician in analyzing large-scale data coming from the longitudinal psychological research project ELSPAC at the Institute for Research on Children, Youth and Family by the Faculty of Social Studies MUNI Brno.
Location
ELG120
Category Conferences - Seminars
In this talk, I will present two topics recently addressed by my students. The first problem falls into audio processing and the second one to image/video processing.

Audio recordings are sometimes affected by defects or even loss of information caused either by obsolete carriers (LP, magnetic tapes) or signal transmission drop-outs. The problem referred to as "audio inpainting" aims at recovering the information in such signal segments. Historically, these problems were solved by interpolation approaches based mostly on autoregressive modelling of partial harmonics. Since most natural signals are sparse with respect to some time-frequency transform, sparse signal priors were utilized in developing related optimization programs. We will compare the old methods with the sparsity-based approach and discuss promising new results based on structured sparsity.

Perfusion MRI is a diagnostic method in medicine used mainly for diagnosing carcinoma and cardiovascular diseases. Herein, a contrast agent is injected in the patient and its concentration then tracked via MRI during time. The signal captured this way from the affected area can be approximated by the lognormal distribution curve. The standard way of obtaining the MR measurements is very slow and does not comply with today’s challenging requirements. We propose using compressed sensing to acquire much less coefficients, having minimal effect on the signal reconstruction. The approach is based on the assumption that the data can be well approxiamted as a sum of low-rank matrix and a matrix sparse in row spectrum.

Links

Practical information

  • Expert
  • Free
  • This event is internal

Organizer

  • Benjamin Ricaud
    LTS2 EPFL

Contact

  • Benjamin Ricaud

Tags

sparsity compressive sensing signal processing audio bioimaging

Event broadcasted in

Share