228
Thesis alerts / Signal Processing
58 (1997) 227-234
Philippe Schroeter*v’ Signal Processing Laboratory
Department of Electrical Engineering, Swiss Federal Institute of Technology, CH-1015 Lausanne, Switzerland
Unsupervised 2-D and 3-D image segmentation2 In this dissertation, different aspects of unsupervised image segmentation are investigated. Algorithms that are robust with respect to noise are developed. Their application over a broad range of images, ranging from gray level images to textured and magnetic resonance images showed accurate segmentation results. In the first part of the thesis, a new spatial clustering scheme based on the fuzzy c-means (FCM), called the constrained fuzzy c-means (CFCM), is elaborated. Image segmentation is obtained by minimization of an objective function composed of the within group sum of square distances (like in the FCM) and of a smoothness constraint resulting from a Gibbs distribution. The minimization problem can be compared to that of a maximum a posteriori possibility, but is much simpler. Moreover, it is shown that the minimum of the objective function of the CFCM with respect to the number of classes corresponds to the “optimum” one. The second part of the dissertation is concerned with the problem of texture image segmentation. In this field, raw images are usually transformed into a set of features corrupted by a large amount of noise, mainly due to modeling errors. For each of
them, a pyramid is built up to a predefined level for reducing the noise in the feature space. The classes and the prototypes are obtained by using the CFCM which further reduces the uncertainty by means of its smoothness constraint. The resolution is gradually restored by projecting down the class labels. We use orientation-adaptive butterflyshaped filters to reduce the class-uncertainty astride the borders. The third part treats the problem of 3-D medical image segmentation. The head is constituted of different types of tissue that can be modeled by a finite mixture of multivariate Gaussian distributions in Magnetic Resonance Images. The goal is to estimate accurately the statistics of desired tissues in presence of other ones of lesser interest. For this purpose, we use a particular genetic algorithm, well suited for estimating the parameters of mixtures of different kind of distributions. These parameters are used in a thresholding-based approach for the automatic 3-D segmentation of the brain. Accurate brain segmentation is obtained by applying 3-D morphological operators on binary masks in an appropriate manner. 3-D visualization of different brains show the quality of the segmentation results.
*Thesis Advisor: Prof. M. Kunt. ‘Currently at: Swiss Telecom PTT, Research and development (FE 226), Guterstrasse 7, CH-3000 Bern 29, Switzerland. E-mail:
[email protected]. ‘A limited number of copies are available upon motivated request.