Skip to Main content Skip to Navigation

Modélisation et apprentissage de relations spatiales pour la reconnaissance et l’interprétation d’images

Abstract : In recent years, the amount of visual data produced by various types of sensors has been continuously increasing. The automatic interpretation and indexation of such data constitute an important challenge in the fields of pattern recognition and computer vision. In this context, the relative position of the different objects of interest depicted in images represents particularly important information for the interpretation of their content. Spatial relations indeed carry rich semantics that are strongly tied with human perception. The research work presented in this thesis thus proposes to explore different generic approaches to the description of spatial information, in order to integrate them in high-level image recognition and understanding systems. First, we present an approach for the description of complex spatial configurations, where objects can be imbricated in each other. This notion is formalized by two novel spatial relations, namely enlacement and interlacement. We propose a model to describe and to visualize these configurations with directional granularity. This model is experimentally validated for applications in biomedical imaging, remote sensing and document image analysis. Then, we present a framework for learning composite spatial relations from image datasets. Inspired by bags of visual features approaches, this strategy allows to build vocabularies of spatial configurations occurring across images, at different scales. These structural features can notably be combined with local descriptions, leading to hybrid and complementary representations. Experimental results obtained for different datasets of structured images highlight the interest of this approach for image recognition and classification tasks.
Document type :
Complete list of metadatas

Cited literature [156 references]  Display  Hide  Download
Contributor : Abes Star :  Contact
Submitted on : Friday, April 26, 2019 - 10:57:22 AM
Last modification on : Saturday, April 11, 2020 - 1:53:55 AM


Version validated by the jury (STAR)


  • HAL Id : tel-02111681, version 1


Michaël Clément. Modélisation et apprentissage de relations spatiales pour la reconnaissance et l’interprétation d’images. Traitement des images [eess.IV]. Université Sorbonne Paris Cité, 2017. Français. ⟨NNT : 2017USPCB024⟩. ⟨tel-02111681⟩



Record views


Files downloads