Date

2014

Author

Karadağ, Özge Öztimur

Metadata

Show full item record

Item Usage Stats

139
views

221
downloads

Cite This

The formulation of image segmentation problem is evolved considerably, from the early years of computer vision in 1970s to these years, in 2010s. While the initial studies offer mostly unsupervised approaches, a great deal of recent studies shift towards the supervised solutions. This is due to the advancements in the cognitive science and its influence on the computer vision research. Also, accelerated availability of computational power enables the researchers to develop complex algorithms. Despite the great effort on the image segmentation research, the state of the art techniques still fall short to satisfy the need of the further processing steps of computer vision. This study is another attempt to generate a “substantially complete” segmentation output for the consumption of object classification, recognition and detection steps. Our approach is to fuse the multiple segmentation outputs in order to achieve the “best” result with respect to a cost function. The proposed approach, called Boosted-MRF, elegantly formulates the segmentation fusion problem as a Markov Random Fields (MRF) model in an unsupervised framework. For this purpose, a set of initial segmentation outputs is obtained and the consensus among the segmentation partitions are formulated in the energy function of the Markov Random Fields model. Finally, minimization of the energy function yields the “best” consensus among the segmentation ensemble. We proceed one step further to improve the performance of the Boosted-MRF by introducing some auxiliary domain information into the segmentation fusion process. This enhanced segmentation fusion method, called the Domain Specific MRF, updates the energy function of the MRF model by the available information which is received from a domain expert. For this purpose, a top-down segmentation method is employed to obtain a set of Domain Specific Segmentation Maps which are incomplete segmentations of a given image. Therefore, in this second segmentation fusion method, in addition to the set of bottom-up segmentation ensemble, we generate ensemble of top-down Domain Specific Segmentation Maps. Based on the bottom–up and top down segmentation ensembles a new MRF energy function is defined. Minimization of this energy function yields the “best” consensus which is consistent with the domain specific information. The experiments performed on various datasets show that the proposed segmentation fusion methods improve the performances of the segmentation outputs in the ensemble measured with various indexes, such as Probabilistic Rand Index, Mutual Information. The Boosted-MRF method is also compared to a popular segmentation fusion method, namely, Best of K. The Boosted-MRF is slightly better than the Best of K method. The suggested Domain Specific-MRF method is applied on a set of outdoor images with vegetation where vegetation information is utilized as domain specific information. A slight improvement in the performance is recorded in this experiment. The method is also applied on remotely sensed dataset of building images, where more advanced domain specific information is available. The segmentation performance is evaluated with a performance measure which is specifically defined to estimate the segmentation performance for building images. In these two experiments with the Domain Specific-MRF method, it is observed that, as long as reliable domain specific information is available, the segmentation performance improves significantly.

Subject Keywords

Image segmentation., Image analysis., Image processing, Markov random fields.

URI

http://etd.lib.metu.edu.tr/upload/12617113/index.pdf
https://hdl.handle.net/11511/23506

Collections

Graduate School of Natural and Applied Sciences, Thesis