Color image segmentation using a novel metric based spectral clustering

This paper lies within the scope of color image segmentation by the spatial-color pixels clustering. It is always difficult to identify the classes of pixels when the color distributions of different scenes highly overlap in the color space and when the color pixels induce non-convex clusters. Spectral clustering has been widely applied to color image segmentation for it can reorganize the pixels representing the same regions into the same group. However, it requires the eigendecomposition of the graph Laplacian matrix G, the computational complexity of which is proportional to O(n³), thus it is not suitable for large data sets. Recently, many approaches have been developed to lower the computational cost. However, many of them generally involve sampling techniques by which a lot of useful information of the original data may be lost. In this work, we design a color image segmentation using an approximate commute time embedding spectral clustering approach. It does not involve any sampling techniques. We apply the designed approach to color images from the Berkeley Segmentation Dataset to verify it. Experimental results show our approach can bring promising results and reduce the runtime significantly.