TAGG

Project title

Multidimensional Classification Models Based On Trainable Feature Aggregation Methods

Funded by

Austrian FWF, P19376-N13

Duration

24 months (01/2008 – 12/2009)

Within this project we want to investigate classification methods that allow the classification of a set of objects. This situation occurs e.g. in image processing, where objects are extracted from the image and a decision (e.g. good/bad) has to be made considering all objects of the image together.

Typical examples are:

• Surface Inspection: A number of faults (= “objects”) on the surface is extracted and each one is represented by a feature vector containing e.g. its size, position, or shape. In order to achieve a good/bad decision, the whole set of faults needs to be considered, e.g. the number of faults, their total area, their spatial distribution or other more complicated aggregated properties.
• Object Recognition: Recognizing objects that are composed of several components. E.g. identifying a house as consisting of an unspecified number of windows and a roof on top of all the windows and some significant edges (the walls) that provide a boundary for the roof and the windows. Or recognizing that a certain group of people lined up in a row is the Brazilian soccer team.
• Biomedical Imaging: The crystallisation patterns of dried biological fluids, in one case native blood drops (clots) on common medical slide, seem to contain a lot of information about diseases and other pathological disorders. Geometrical and colour features are extracted out of scanned images from these clots, but only the accumulation of all these varying features of different regions of the blood spot decides, whether the patient has some pathological disorder.

 

 The common requirements of these classification tasks are

• the number of objects in the set is different for each image
• the classification task requires aggregated information from the whole set of objects
• the classifier has to be trainable, since the classification rules are not known beforehand.

In image processing this combination of requirements has not yet been investigated and is not covered by existing classification methods as they are usually focused on classifying only single feature vectors. The goal is to develop aggregation methods that bring us a little bit closer to a model of human decision making.