Reconnaissance de documents assistée : architecture logicielle et intégration de savoir-faire

Frédéric Bapst

Abstract

This thesis is devoted to document recognition with an assisted perspective, advocating an adequate combination between human and machine capabilities. Our contributions tackle various aspects of the underlying software architecture. We outline the recognition systems of the future, focusing on human-machine dialog and incremental learning opportunities.

In our architecture, data are organized in a modular and homogeneous way. The control is decentralized according to a multi-agent modeling. The implementation platform is based on concurrent, distributed, and multi-languages programming. Our general architecture, demonstrated through the current prototype, can serve as the basis of a wide variety of applications.

Our software architecture takes advantage of the typographical know-how, through the use of a standardized font management support. This integrated approach lets us enhance the ergonomy, extend the possible use of the recognition results, and redefine some recognition techniques. The study presents some innovative algorithms, developments, and experiments, in the field of optical character recognition, font identification, or segmentation.

Besides, we bring a contribution to the problem of measuring the performance of cooperative recognition systems, through the introduction of a new cost model. Our notations are able to describe assisted recognition scenarios, where the user takes part in the process, and where the accuracy is modified dynamically thank to incremental learning. Our cost model is illustrated both in simulations and in experiments.

Keywords

This thesis is part of the CIDRE project (Cooperative & Interactive Document Reverse Engineering).
SDE Group Home Page
Last updated September 7, 1998