List of Figures

1. Normalized number of strong attractors in the range domain with respect to the relative noise level $\varepsilon$. The results are given for two values of fixed bandwidth.
2. Segmentation for different $h_0$. In images (g) and (h) segments smaller than 50 pixels are connected to the neighbor with the most similar color. Region boundaries are superimposed to emphasize details.
3. Comparison of the FHS with other methods. The output differs for different algorithms. The EGB uses random coloring scheme. JSEG algorithm superimposes region boundaries on the original image.
4. $NPR$ index varying algorithm parameters: (a) maximal $NPR$ index, (b) distribution of the maximal $NPR$ index, (c) mean $NPR$ index, (d) standard deviation of the $NPR$ index
5. Algorithms running time: Plots (a), (b) and (c) show running times of evaluated algorithms vs. controlling parameters. Plot (d) represents running time vs. image resolution.

Figure 2: Segmentation for different $h_0$. In images (g) and (h) segments smaller than 50 pixels are connected to the neighbor with the most similar color. Region boundaries are superimposed to emphasize details.

Original image	Original Image
$h_0 = 8$	$h_0 = 8$
$h_0 = 16$	$h_0 = 16$
$h_0 = 16$, min. size = 50pix	$h_0 = 16$, min. size = 50pix

Figure 3: Comparison of the FHS with other methods. The output differs for different algorithms. The EGB uses random coloring scheme. JSEG algorithm superimposes region boundaries on the original image.

Original image	MS EDISON
FHS, $h_0 = 4$	EGB
FHS, $h_0 = 12$	JSEG
FHS, $h_0 = 20$	GNM

Figure 4: $NPR$ index varying algorithm parameters: (a) maximal $NPR$ index, (b) distribution of the maximal $NPR$ index, (c) mean $NPR$ index, (d) standard deviation of the $NPR$ index

Figure 5: Algorithms running time: Plots (a), (b) and (c) show running times of evaluated algorithms vs. controlling parameters. Plot (d) represents running time vs. image resolution.