Authors:
Lassi Korhonen
and
Keijo Ruotsalainen
Affiliation:
University of Oulu, Finland
Keyword(s):
Spectral Clustering, Image Segmentation, Diffusion.
Related
Ontology
Subjects/Areas/Topics:
Image and Video Processing, Compression and Segmentation
;
Multimedia
;
Multimedia Signal Processing
;
Telecommunications
Abstract:
Different clustering algorithms are widely used for image segmentation. In recent years, spectral clustering has risen among the most popular methods in the field of clustering and has also been included in many image segmentation algorithms. However, the classical spectral clustering algorithms have their own weaknesses, which affect directly to the accuracy of the data partitioning. In this paper, a novel clustering method, that overcomes some of these problems, is proposed. The method is based on tracking the time evolution of the connections between data points inside each cluster separately. This enables the algorithm proposed to perform well also in the case when the clusters have different inner geometries. In addition to that, this method suits especially well for image segmentation using the color and texture information extracted from small regions called patches around each pixel. The nature of the algorithm allows to join the segmentation results reliably from different s
ources. The color image segmentation algorithm proposed in this paper takes advantage from this property by segmenting the same image several times with different pixel alignments and joining the results. The performance of our algorithm can be seen from the results provided at the end of this paper.
(More)