Due to the large amount of multimedia documents either already existing or produced daily, we are faced with the problem of retrieving information from gigantic databases, for which a purely manual indexing process is currently impossible. In this context, it has become necessary to design new techniques, if not to extract the whole semantic content of a given document, but at least to produce a first structure of it automatically.
Dealing only with video documents, this thesis therefore proposes to build automatic tools that create such a structure in two steps. As a result of the first linear part of the structuring, the video document is splitted into different entities that go from the scene to the object, through the shot, the part of shot and the key frame. The second and relational step consists in extracting relationships, by establishing all kinds of semantic and syntactic links, between the different types of entities. In addition to being general and automatic, the tools proposed were elaborated with respect to a precise methodology. According to this methodology we only use simple and low level image processing criteria, and in particular these from mathematical morphology. These criteria, when combined together and also with logical rules of decision, already enable us to obtain a structure which is consistent, efficient and representative of a content with a high semantic level. The choice of this methodology also leads to a very high speed, as our tools work faster than real time. These tools are validated through numerous examples and applications, based mainly on television news broadcasts documents.
Indexing, Image and sequence processing, Mathematical morphology, Colour segmentation , Video segmentation and structuring, Shot relation extraction, Semantic content.
Last Update: 27 - 03 - 00
Claire-Hélène Demarty's Home Page
CMM Home Page