13 sep
13/09/2018 14:00

Sciences & Société

Soutenance de thèse : Louisa KESSI

Unsupervised Recognition based on spatial relationships in High Dimensional space for colored business documents structures and Object Recognition

Doctorante : Louisa KESSI

Laboratoire INSA : LIRIS
Ecole doctorale : ED512 INFOMATHS

This thesis have the objective to develop innovative system to recognize automatically the logical structures of digitized business documents having different layouts and various contents, without an explicit model. The existing commercial software which read automatically the structure of invoices are all rule-based systems, which use templates of marks in order to localize the zones of interest associated with a logical label. Because each company designs their own invoices, it is difficult to manage a rule-based system or to train a recognition system for all existing business documents layout. As a model- driven approach is impossible to use for all possible documents layout, we propose unsupervised system that are able to process all types of documents structures. There is two difficult problems to overcome: the recognition of logical structure by using the spatial relations between objects and marks and the large variety of possible layouts to recognize. That is why we have validate our developed theory on computer vision applications like objects recognition on classical database like Pascal VOC or ImageNet. We developed the first color segmentation software called AColDPS for noisy and degraded color documents. AColDPS automatically segments all documents without any prior information and separates the different colors of characters by using color morphological and geodesic operations. The software separates the texts from the graphic elements by using colors features even if the elements are overlapped or by geodesy if they are of the same color. The high level stage of the system consists on the modelization of spatial relation, part-based recognition, high dimensional feature space, distances in such space and unsupervised approaches. Logical structure recognition is a very complex problem because the recognition of each part of a structure depends simultaneously of the recognition of the neighboring parts and the spatial relation between the different parts.