Two methods of random seed generation to avoid over-segmentation with stochastic watershed: application to nuclear...

A stochastic version of the watershed algorithm is obtained by choosing randomly in the image the seeds from which the watershed regions are grown. The output of the procedure is a probability density function corresponding to the probability that each pixel belongs to a boundary. In the present paper, two stochastic seed-generation processes are explored to avoid over-segmentation. The first is a non-uniform Poisson process, the density of which is optimized on the basis of opening granulometry. The second process positions the seeds randomly within disks centred on the maxima of a distance map. The two methods are applied to characterize the grain structure of nuclear fuel pellets. Estimators are proposed for the total edge length and grain number per unit area, L(A) and N(A), which take advantage of the probabilistic nature of the probability density function and do not require segmentation.