This thesis tackles the problem of technical document interpretation applied to France Telecom Documentation. This subject is on the crossroad of different fields like signal or image processing, pattern recognition, artificial intelligence, man-machine interaction and knowledge engineering. Indeed, each of these different fields can contribute to build a reliable and efficient document interpretation device. In this interdisciplinary context, this thesis is divided in two main parts. The first part is considering an original method used to detect and recognise multi-scaled and multi-oriented patterns like symbols or characters. The theoretical basis of this method is given by the Fourier-Mellin transform. It allows recognising isolated patterns but also, in some cases, connected patterns. The approach also allows the estimation of shape's movement parameters. Tools that have been developed in this context are evaluated regarding the state of the art in optical characters recognition. Obtained results with this original method are really competitive. The second part is focusing the theme of technical document analysis under the point of view of knowledge engineering. The aim is to show the feasibility and relevance of a "knowledge based approach" in the context of technical document interpretation. An external and explicit knowledge model, a distributed agent-based software architecture and several user interfaces give the main concepts of this approach. A first implementation using these concepts is shown through a presentation of a system named "NATALI v2". This implementation has good reliability and adaptability properties.
