Two proposals for k-means-based techniques for fully automatic image segmentation

The aim of this paper is to suggest two possible alternatives to create fully automatic adaptive k-means segmentation algorithms. As is usual in the literature we model the gray scale values of the volume image with a White Gaussian Process and we superimpose a prior model on the region process in the form of Markov Random Field. These assumptions require the use ofestimators for the parameters of the two functions. The proposed estimation ofthe mean and the standard deviation of the White Gaussian Process is adaptive, and this has been achieved using rectangular windows of decreasing size whichdelimits the region where the mean has to be computed. The initial estimates are computed over the whole image and then it is refineddecreasing the size of the window. The Hammersley-Clifford theoremallows us to model the region process in term of a Gibbs Distribution. The Gibbs parameter βis estimated using a correlation-based technique. The segmentation is obtained maximizing the a poster density function (MAP estimation) using an Iterated Conditional Modes (ICM) technique.Unlike other techniques the proposed techniques are fully automatic, in the sense that all the parameters of the model are estimated during the segmentation process, and there is no need to fix a priori values that, as we will prove, can effect the outcome of the segmentation