Related work

Although the research in this area has led to many different image segmentation techniques presented in the literature [6], no single algorithm can be considered good for all applications and all images [7]. Segmentation methods used for color images can be divided into two main categories: feature-space based techniques (clustering methods [8,9,10] and histogram multi-thresholding [7,11,12]) and image-domain based techniques. The latter are further divided into pixel-similarity based algorithms [13,14,15] and pixel-difference based algorithms [2,4,16]. Image-domain based techniques exploit the pixel context interaction which increase computational complexity. The most common feature-space techniques gained their popularity through the adaptive k-means algorithm [17,18]. Most of these methods are parameterized with number of expected clusters. Number of clusters is determined automatically in the Mean Shift algorithm [1] using nonparametric density estimation and gradient guided hill-climbing procedure.

Histogram-based approaches relay on the estimation of the discrete density in the designated color space where the clusters are represented by hills in the histogram. These methods are not parameterized with the a priori knowledge about number of clusters or their shape. The result depends on the discretization resolution of the feature space and the quality of provided segmentations can degrade if scale variations are present in the input data. To overcome this drawback, the Hill-manipulation algorithm [19] adaptively refines histogram resolution in the regions of higher density. Beside abovementioned, methods exist that combine information contained in different attributes or apply postprocessing to improve the quality of the segmentation acquired by color quantization [20,21].