Schuster, Hanns-Florian: Interpretation of Aerial Images with Learned Graphical Models. - Bonn, 2016. - Dissertation, Rheinische Friedrich-Wilhelms-Universität Bonn.
Online-Ausgabe in bonndoc: https://nbn-resolving.org/urn:nbn:de:hbz:5n-42838
@phdthesis{handle:20.500.11811/6600,
urn: https://nbn-resolving.org/urn:nbn:de:hbz:5n-42838,
author = {{Hanns-Florian Schuster}},
title = {Interpretation of Aerial Images with Learned Graphical Models},
school = {Rheinische Friedrich-Wilhelms-Universität Bonn},
year = 2016,
month = feb,

note = {In this thesis we present a new approach for image interpretation of aerial images using learned graphical models.
The approach uses a region based hierarchical feature adjacency graph. This contains homogeneous regions that were extracted out of the image as well as its neighbor relationships.
For each region there are 17 image features extracted to describe the region, e.g. color, structure and symmetries. The neighbor relations are attributed by seven features describing their geometrical relation. The Bayes net consists of nodes for regions and their image features as well as the neighbor relationships. It also models the higher aggregated elements with nodes for cliques, objects and the image scene. The regions of this adjacency graph and the describing features are used as observations for the nodes of the Bayes net. In the first learning step also the scene node is introduced as observation the other nodes are hidden i.e. not observed.
The presented approach has two stages. In the first stage the Bayes net is trained with known ground truth data. By introducing the observations, a structural learning algorithm searches the best net structure and learns the probability distributions and the dependencies of the nodes of the Bayes net. The learned parameters are the result of the first stage. They are saved and used in the second stage.
The second stage interprets new images using a Bayes net with the parameters of the first stage. Again, the regions and features of the region based feature adjacency graph are introduced as observations. Using an iterative maximum a posteriori estimation, we search for the optimal Bayes net structure to describe the underlying image. The states of the nodes of the Bayes net represent now the interpretation according to our learned vocabulary.},

url = {https://hdl.handle.net/20.500.11811/6600}
}

Die folgenden Nutzungsbestimmungen sind mit dieser Ressource verbunden:

InCopyright