Effects of color spaces and distance norms on graph-based image segmentation

Use of the graph theory tools in image processing field is growing up with each passing day. Graph theory makes the operations easier for image processing applications, and can represent digital image components completely. In image segmentation processes, the graph theory tools are also used widely. These kinds of image segmentation processes are called graph-based image segmentation. In many image processing applications, it seems as a problem that which color space the color values of pixels should be considered according to and which distance norm should be used to measure the difference between two points in the space. In this work, a graph-based image segmentation algorithm is tested on several color spaces with different distance norms. The test is carried out on 100 real world images that take part in a general-purposed image segmentation dataset. The average segmentation results are given as F-measure in this work with regard to both color spaces and distance norms. The results show that L*a*b* and L*u*v* color spaces are more appropriate than RGB color space, in general. The squared Euclidean distance norm gives more accurate results than the Euclidean distance norm, used in the source paper, if the Gaussian smoothing is not used as pre-processing.