Friday, 14 December 2018

MARCH : Multi-classifier systems (MCS)

From LISA

Jump to: navigation, search

Contents

Description

In numerous pattern recognition or image analysis applications, it is becoming essential to develop automatic means for transforming spectral image values into map classes relevant to the applications concerned. For instance, every pixel of an image can be characterized as belonging to a given region (a class being then a set of regions identically labeled but distributed in the image) so that, at the end, a rough image given in input is divided into clearly identified regions in output. This process which, on the basis of the pixel attributes (provided by image processing techniques applied on the pixel and its neighborhood), automatically assigns a ``region label to this pixel can be seen as a classification task. Whenever some formal or analytical way or some explicit prior knowledge to provide the pixel with its label is lacking, the most convenient alternative is to resort to data-driven or classification methods, which, on the basis of an important and reliable data set, can automatically discover the classification laws. These laws are expected to adequately apply to new and unlabeled data. This is all what classification methods are about: finding the right data and attributes, finding the right parameterized structure to capture the laws, finding the right cost to minimize and the right associated or optimization algorithm, so that, finally, the resulting classification laws are as closed as possible to the theoretical but unknown laws and hopefully robust enough to correctly identify all new data to come. This helps to explain the recent importance gained by methods such as nearest neighbors, decision tree, neural nets, fuzzy systems or genetic algorithms for mapping the lowest and primitive level of image analysis to a first more abstract and meaningful level.

Applications as the ones just described present further peculiarities that allow to enrich classical classification methods with aspects or constraints characteristic of images, and so doing to lead to the development of original methodologies at the border of image analysis and data processing.

Four aspects resulting from the interaction of these two fields will be investigated in the project.

  • Hierarcical Classification
  • Combining Classifiers
  • Top-down Revision
  • Double Supervision

(Begin of the project 1998)

Teams

IRIDIA Artificial Intelligence Institute

Professor Hugues Bersini

Researchers :

  • Mauro Birattari
  • Patrice Latinne

SLN Logical and Digital Systems

Professor Philippe Van Ham

Researchers :

  • Olivier Debeir
  • Thierry Leloup

SMA Department of Applied Mechanics

Professor Alain Delchambre

Researchers :

  • Emanuel Falkenauer
  • NathanaĆ«l Ackerman
  • Olivier Gresse

Collaborators

  • Eleonore Wolff, IGEAT, geographical information systems and remote-sensing (ULB).
  • Christine Decaestecker, laboratory of histology, Faculty of Medecine (ULB).
  • Marco Saerens, CSC Benelux.



Related publications

[1]

Link

Link to the project home page [2]