Parallel Image Segmentation using Fiedler Vector

Bora Akaydin and Murat Manguo─člu


Image Segmentation, Parallel Computing, Fiedler Vector, Normalized Cut


In this work, we propose a parallel algorithm for image segmentation. Our method is an unsupervised image segmentation procedure, which is a combination of a graph based image segmentation algorithm and parallel Fiedler vector computation algorithm. In this work, whole image is handled as a weighted graph, whose vertices are the pixels. As the segmentation method, we use Fiedler vector, which is the eigenvector which corresponds to the second smallest eigenvalue of the graph Laplacian matrix. Laplacian matrices belonging to big images can be undesirably large, thus the Laplacian matrix of an M x N sized image has the size of (M x N) x (M x N). Parallelization is needed since the calculation of Fiedler vector with a serial algorithm is time consuming and the Laplacian matrix is large. The visual results of segmentation process is satisfactory and also scale-up for parallelization is well for at least 64 processors.

Important Links:

Go Back