E-Locus - Institutional Repository of the University of Crete - Bayesian flooding for image and video segmentation

Home Bayesian flooding for image and video segmentation

Results - Details

[Add to Basket]

Identifier

000342575

Title

Bayesian flooding for image and video segmentation

Alternative Title

Αλγόριθμοι Πλημμυρίδας κατά Bayes για Τμηματοποίηση Εικόνων και Βίντεο

Author

Γκρίνιας, Ηλίας

Abstract

Image segmentation is one of the fundamental problems in image processing and computer vision. Segmentation is also one of the first steps in many image analysis tasks. Image understanding systems such as face or object recognition often assume that the objects of interest are well segmented. Different visual cues, such as colour and texture in still images and motion in image sequences, help in achieving segmentation. Segmentation is also goal dependent, subjective, and hence ill-posed in a general set up. However, it is desirable to consider generic methods that can be applied to a large variety of images and can be adapted for specific applications. This thesis work focuses on developing such segmentation methods that work on natural images.
Segmentation is based on a statistical framework. Visual feature description is unified under a statistical point of view. The first part of the framework proposes a new, block based clustering method for visual content classification. Starting with the computation of visual cues, statistics of blocks are estimated and a k-means algorithm is employed for classification of blocks in a number of classes, using statistical dissimilarity criteria. A novel distance metric between affine models of optical flow is also described. Towards the automation of clustering, a method for computing the number of visual classes as well as a feature selection procedure are also proposed.
Second part of the framework explores region based segmentation, given the statistical description of classes. Initial regions of high confidence per class are determined and two new region growing algorithms are proposed to expand initial regions. The strong relation of region growing on statistical surfaces to the connectivity percolation process is also underlined.
To test the effectiveness of these new techniques, extensive tests are conducted on the Berkeley segmentation data set and the associated ground truth, using colour and texture. Furthermore, segmentation of independently moving objects using interframe difference and colour as well optical flow based segmentation in image sequences is also described. Finally, an application is shown, in which the proposed framework is used for extracting left ventricle in medical cardiac images.

Language

English

Issue date

2009-04-02

Collection

School/Department--School of Sciences and Engineering--Department of Computer Science--Doctoral theses

Type of Work--Doctoral theses

Views

532

Digital Documents
	Download document View document Views : 20