An efficient curve evolution algorithm for multiphase image segmentation

Published

January 7, 2015

Author(s)

Gunay Dogan

Abstract

We propose a novel iterative algorithm for segmentation of multiphase images by curve evolution. Specifically, we address a multiphase version of the Chan-Vese piecewise constant segmentation energy. Our algorithm is efficient: it is based on an explicit Lagrangian representation of the curves and it converges in a relatively small number of iterations. We devise a stable curvature-free semi-implicit velocity computation scheme. This enables us to take large steps to achieve sharp decreases in the multiphase segmentation energy when possible. The velocity and curve computations are linear with respect to the number of nodes on the curves, thanks to a finite element discretization of the curve and the gradient descent equations, yielding essentially tridiagonal linear systems. The step size at each iteration is selected using a nonmonotone line search algorithm ensuring rapid progress and convergence. Thus, the user does not need to specify fixed step sizes or iteration numbers. We also introduce a novel dynamic stopping criterion, robust to various imaging conditions, to decide when to stop the iterations. Our implementation can handle topological changes of curves, such as merging and splitting as well. This is a distinct advantage of our approach, because we do not need to know the number of phases in advance. The curves can merge and split during the evolution to detect the correct regions, especially the number of phases.

Volume

8932

Conference Dates

January 13-16, 2015

Conference Location

Hong Kong

Conference Title

10th International Conference on Energy Minimization Methods in Computer Vision and Pattern Recognition

Pub Type

Conferences

Download Paper

DOI Link

Keywords

Multiphase image segmentation, shape optimization

Citation

Dogan, G. (2015), An efficient curve evolution algorithm for multiphase image segmentation, 10th International Conference on Energy Minimization Methods in Computer Vision and Pattern Recognition, Hong Kong, -1, [online], https://doi.org/10.1007/978-3-319-14612-6_22 (Accessed April 18, 2024)