Skip to Main content Skip to Navigation

Modèles graphiques probabilistes pour la reconnaissance de formes

Sabine Barrat 1
1 QGAR - Querying Graphics through Analysis and Recognition
LORIA - Laboratoire Lorrain de Recherche en Informatique et ses Applications
Abstract : The rapid growth of Internet and multimedia information has shown a need in the development of multimedia information retrieval techniques, especially in image retrieval. We can distinguish two main trends. The first one, called "text-based image retrieval", consists in applying text-retrieval techniques from fully annotated images. The text describes high-level concepts but this technique presents some drawbacks: it requires a tedious work of annotation. Moreover, annotations could be ambiguous because two users can use different keywords to describe a same image. Consequently, some approaches have proposed to use Wordnet in order to reduce these potential ambiguities. The second approach, called "content-based image retrieval", is a younger field. These methods rely on visual features (color, texture or shape) computed automatically, and retrieve images using a similarity measure. However, the obtained performances are not really acceptable, except in the case of well-focused corpus. In order to improve the recognition, a solution consists in combining differents sources of information: for example different visual features and/or visual and semantic information. In many vision problems, instead of having fully annotated training data, it is easier to obtain just a subset of data with annotations, because it is less restrictive for the user. In this direction, this thesis deals with modeling, classifying, and annotating images. We present a scheme for natural image classification optimization, using a joint visual-text clustering approach and automatically extending image annotations. Moreover, we propose a symbol recognition method, by combining visual features. The proposed approach is derived from the probabilistic graphical model theory and dedicated for both tasks of weakly-annotated image classification and annotation. We consider an image as weakly annotated if the number of keywords defined for it is less than the maximum defined in the ground truth. Thanks to their ability to manage missing values and to represent possible relations between keywords, probabilistic graphical models have been proposed to represent weakly annotated images. Therefore, the proposed model does not require that all images be annotated: when an image is weakly annotated, the missing keywords are considered as missing values. Besides, our model can automatically extend existing annotations to weakly-annotated images, without user intervention. The uncertainty around the association between a set of keywords and an image is tackled by a joint probability distribution over the dictionary of keywords and the visual features extracted from our collections of images. Our model is also used to recognize symbols by combining different kinds of visual features (continuous and discrete features). Moreover, in order to solve the dimensionality problem due to the large dimensions of visual features, we have adapted a variable selection method. Finally, a system of image retrieval has been proposed. This system enables keyword and/or image requests and relevance feedback process. The experimental results, obtained on large image databases (general or specialized databases), show the interest of our approach. Finally, the proposed method is competitive with a state-of-art model.
Document type :
Complete list of metadatas

Cited literature [236 references]  Display  Hide  Download
Contributor : Sabine Barrat <>
Submitted on : Friday, October 29, 2010 - 6:13:17 PM
Last modification on : Tuesday, April 24, 2018 - 1:54:42 PM
Long-term archiving on: : Sunday, January 30, 2011 - 3:07:13 AM


  • HAL Id : tel-00530755, version 1



Sabine Barrat. Modèles graphiques probabilistes pour la reconnaissance de formes. Interface homme-machine [cs.HC]. Université Nancy II, 2009. Français. ⟨tel-00530755⟩



Record views


Files downloads