Section: New Results

Around the Kasimir research project

Participants : Nicolas Jay, Jean Lieber, Bart Lamiroy, Amedeo Napoli, Thomas Meilender.

This special research project involves researchers working around the Kasimir project and Bart Lamiroy who was attached to the Orpailleur Team during his “INRIA délégation” (2010-2011) and at the same time was a visiting scientist at Lehigh University, USA. The background of Bart Lamiroy is in document and image analysis. Recently he was interested in investigating the application of KDDK to numerical and structural data including document images. The objective is to extend mining tools towards complex and semi-structured multi-media data on the one hand, and to associate image analysis with KDDK techniques on the other hand.

The main research direction which is followed at the moment is in concern with the Kasimir project. Actually, oncology protocols are mainly documented and represented in diagram formats. The classification and CBR techniques used in the Kasimir project require that the ontologies and decision protocols have to be represented in OWL. Based on previous work, we started modeling the mapping of visual features in diagram charts with semantics of the medical domain ontology. The mapping between the visual ontology and the domain ontology should guide a more complete extraction of the protocols from the diagrams for completing the domain ontology of the Kasimir system.

Moreover, during his stay at Lehigh University, Bart Lamiroy developed a new approach for recovering useful information within image data. By recording a wide range of “provenance information” related to complex image analysis processes, the DAE platform (http://dae.cse.lehigh.edu ) provides a large set of metadata that can be used by KDDK methods. For example, this allows the correlation and combination of numerical and symbolic aspects, e.g. relating image aspects and domain symbolic representations (within domain ontologies). This work bridges the gap between formal knowledge representation and signal-based pattern recognition and offers a robust experimental environment for further application of KDDK on image data.