Seminars – 2019-2020

The seminars of OBELIX team are currently held every two weeks at the IRISA lab, Tohannic campus, room D106.

The seminar is coordinated by Chloé Friguet.

Previous seminars – 2019-2020

  • Date: January, 6 (Monday) – 14h00
  • Room : D106
  • Speaker: Arthur Pesah (Machine Learning Scientist, 1QBit, Toronto, Canada)
  • Title:  Quantum Machine Learning Beyond the Hype
  • Abstract: While it is hard to find an expression containing more buzz words than “quantum machine learning”, the emergence of this field has been driven by a real scientific fact: quantum computers are fast linear algebra solvers, and machine learning is mostly about solving linear algebra problems. This simple fact has led to many different ideas of connections between the two fields: can we solve SVM, PCA, recommender systems or clustering exponentially faster with the help of a quantum device? Could variational quantum circuits have any advantage over deep neural network? Are there optimization problems that could benefit from having a quantum subroutine?In this talk, I will start by introducing the basics of quantum computing and the current practical state of those devices. I will then discuss their potential to help improve or accelerate machine learning algorithms, as well as the current challenges of the field.

  • Date: November, 25 (monday, 9h30)
  • Room : Amphi. Yves Coppens
  • Speaker: Jamila MIFDAL (IRISA Obelix and LMBA, Univ. Bretagne-Sud)
  • Title: Application of optimal transport and non local methods to hyperspectral and multispectral image fusion (PhD defense)
  • Abstract: The world we live in is constantly under observation. Many areas such as offshore zones, deserts, agricultural land and cities are monitored. This monitoring is done throughout remote sensing satellites or cameras mounted on aircrafts. However, because of many technological and financial constraints, the development of imaging sensors with high accuracy is limited. Therefore, solutions such as multi-sensor data fusion overcome the different limitations and produce images with high quality.This thesis is about hyperspectral and multispectral image fusion. A hyperspectral image (HS) has a high spectral resolution and a low spatial resolution, whereas a multispectral image (MS) has a high spatial resolution and a low spectral resolution. The goal is the combination of the relevant information contained in each image into one final high resolution one. In this dissertation various methods for dealing with hyperspectral and multispectral image fusion are presented. The first part of the thesis uses tools from the optimal transport theory namely the regularized Wasserstein distances. The fusion problem is thus modeled as the minimization of the sum of two regularized Wasserstein distances. In the second part of this thesis, the hyperspectral and the multispectral fusion problem is presented differently. The latter is modeled as the minimization of four energy terms including a non local term. Experiments were conducted on multiple datasets and the fusion was assessed visually and quantitatively for both fusion techniques.The performance of both models compares favorably with the state-of-the-art methods.

  • Date: November, 25 (monday, 14h30)
  • Room : salle B133 bât. Yves Coppens (LMBA)
  • Speaker: Julie Delon (Univ. Paris Descartes, IUF)
  • Title: Transport optimal entre mélanges de gaussiennes
  • Abstract: Les modèles de mélanges de gaussiennes sont très utilisés en statistique, ces modèles s’avérant notamment utiles lorsque l’on souhaite représenter des données réelles. Le transport optimal peut servir à calculer des distances entre de tels mélanges ou à les interpoler, mais les barycentres ainsi obtenus ne conservent généralement pas la propriété d’être un mélange de gaussiennes. Dans cet exposé, nous introduirons une distance de type Wasserstein définie en restreignant l’ensemble des mesures de couplage à des mélanges de gaussiennes. On dérivera une formulation discrète très simple de la distance correspondante, formulation qui la rend bien adaptée aux problèmes en grande dimension. Nous étudierons également la formulation multimarginales du même problème. L’exposé sera illustré par des exemples d’applications en traitement ou édition d’images.

  • Date: November, 14
  • Room : D106
  • Speaker: Anne OSIO (Technical University of Kenya)
  • Title: Monitoring Riparian Vegetation Degradation using Sentinel-1 and Sentinel-2
  • Abstract: The study proposes a workflow for detection and classification of degrading acacia vegetation along Lake Nakuru riparian reserve. Dual polarized Sentinel-1 inhibits the VV+VH bands which are sensitive to soil moisture, woody and herbaceous vegetation. The polanimetric properties of the Synthetic Aperture Radar (SAR) C-band were used to discriminate the following Land cover classes: Acacia forest, Degraded Forest, Degraded-submerged forest, and water. A set of six seasonal corresponding S1 and S2 images (2018-2019) were pre-processed and used in  the classification and hence discrimination of the four land cover classes. Previous studies on the use of cross polarized VV+VH Intensity bands  on vegetation has shown that there is a correlation between vegetation indices derived from optical sensors and the backscatter indices derived from the same Land cover classes.

Comments are closed.