Interactive image segmentation based on the Potts model

Angelopoulos Vasileios

Semester Project

Communication Systems Section, EPFL

June 2015

Abstract

This project was based on a mathematical segmentation model, which is formed as a minimization problem, called Potts model. Using the ADMM strategy the multidimensional Potts problem, which is an NP-hard one, is split into univariate Potts problems that are solvable exactly and quickly. Using and explaining the mathematical theory of this model, we created an Icy platform plugin, which gives the opportunity to the user to segment an image in full or in preview size. The latter produces a preview result using a rescaled version of the input image. Using this process, we give the result to the user faster. Moreover, the supervised part of the application includes merging of two or more segments, or refinement of an existing segment. There, one, that is not satisfied by the initial result, can modify the „ parameter for this segment in order to achieve more refined segmentation. Finally, the output, after using all of the functions of the application, is given in 3 different ways:

segmentation with mean values of the pixels included to each of the segments,
boundaries of the segmentation and
HSV to RGB transformation of the result combining information of the input image, its segmentation and the boundaries of the latter.