Color Image Segmentation: Kernel Do the Feature Space

Abstract

In this paper, we try to apply kernel methods to solve the problem of color image segmentation, which is attracting more and more attention recently as color images provide more information than gray level images do. One natural way for color image segmentation is to do pixels clustering in color space. GMM has been applied for this task. However, practice has shown that GMM doesn’t perform this task well in original color space. Our basic idea is to solve the segmentation in a nonlinear feature space obtained by kernel methods. The scheme is that we propose an extension of EM algorithm for GMM by involving one kernel feature extraction step, which is called K-EM. With the technique based on Monte Carlo sampling and mapping, K-EM not only speeds up kernel step, but also automatically extracts good features for clustering in a nonlinear way. Experiments show that the proposed algorithm has satisfactory performance. The contribution of this paper could be summarized into two points: one is that we introduced kernel methods to solve real computer vision problem, the other is that we proposed an efficient scheme for kernel methods applied in large scale problems.