Centre de documentation
scientifique

Kontextsensitive Bildinterpretation mit Markoff-Zufallsfeldern / M. Köster (1995)

Public Titre : Kontextsensitive Bildinterpretation mit Markoff-Zufallsfeldern Titre original : [Photo-interprétation en tenant compte du contexte avec les champs aléatoires de Markov] Type de document : Thèse/HDR Auteurs : M. Köster, Auteur Editeur : Munich : Bayerische Akademie der Wissenschaften Année de publication : 1995 Collection : DGK - C Sous-collection : Dissertationen num. 444 Importance : 73 p. Format : 21 x 30 cm ISBN/ISSN/EAN : 978-3-7696-9487-1 Note générale : Bibliographie Langues : Allemand (ger) Descripteur : [Vedettes matières IGN] Traitement d'image [Termes IGN] champ aléatoire de Markov [Termes IGN] classification bayesienne [Termes IGN] distribution de Gibbs [Termes IGN] image aérienne [Termes IGN] optimisation (mathématiques) [Termes IGN] reconnaissance d'objets [Termes IGN] reconnaissance de formes [Termes IGN] segmentation d'image [Termes IGN] théorème de Bayes [Termes IGN] voisinage (relation topologique) Index. décimale : 35.20 Traitement d'image Résumé : (Auteur)This thesis solves the problem of interpretation of digital images by Markov random fields in connec-tion with Bayesian statistics. The interpretation requires the description of the objects expected in an image and of the objects existing in the image. The appearence of objects in images is ambiguous, so that context relations between neighbouring objects have to be included. Object recognition requires the labeling of segmented image primitives. By this process uncertainties appear which have to be considered by the object description as well as by the labeling. A mathematical, stochastic method is used, which is theoretically founded and domain independent. Based on Markov random fields a relational description is created for the expected and the existing objects. The relations between neighbouring objects are represented by potentials of cliques of the Gibbs distributions of the Markov random fields. This creates a context dependency based on the local neighbourhood. The flexibility of Bayesian statistics allows the introduction of prior informa-tion of the unknown object labels, which consists of the frequencies of the occurrence of objects and their necessary neighbourhood relations. The labeling problem is solved by maximizing the posteriori density which follows from the Bayes' theorem. Deterministic and stochastic optimization algorithms are compared. The interpretation model not only contains features of the objects and relations {o their neighbours, indirect neighbourhood objects and three pairwise neighbouring objects are also included. An inter--pretation is therefore possible with a more detailed consideration of object relations. A semantic model of concrete objects and their context relations in aerial images is constructed. If there are no images of the same domain, elements in the object model will be determined by heuristic arguments. The approach has been tested on a generated and a real image. The ideal circumstances of the synthe-tic image are used, to investigate influences of variations of the image model and of the optimization algorithms. Finally a segmented image with three color channels is interpreted. Numéro de notice : 28045 Affiliation des auteurs : non IGN Thématique : IMAGERIE Nature : Thèse étrangère Permalink : https://documentation.ensg.eu/index.php?lvl=notice_display&id=63392

Exemplaires(2)

Code-barres	Cote	Support	Localisation	Section	Disponibilité
28045-01	35.20	Livre	Centre de documentation	Télédétection	Disponible
28045-02	35.20	Livre	Centre de documentation	Télédétection	Disponible