Often image captured in real world contains objects that
are constructed with tiny lines. An image from the Berkeley
Segmentation Dataset (BSDS500) [10] contains the steel
roller coaster tracks is tested for the segmentation and its
result is shown in Fig. 9. As it can be seen that for the 6 × 6
image cells segmentation, the part marked with red circle
ring on the top left of the clustered segment 2 has a sharp
right angle edge. This is due to normalised cuts algorithm is
not run on the corresponded cell. However, for the 5 × 5
image cells segmentation, the sharp corner edge does not
occur. The edges of the segment follow closely with the object boundaries. When the number of image cells
increased on the same image, the area covered by each of the
image cells has a zoom in effect on it. This decreases the
discrimination power for the area. In the first cell (top left) of
5 × 5 image cells, there is a critical edge boundary that is
supposed to be identified. However, the edge does not take
up much of the area in the cell. The normalised cuts
algorithm omits the cell that required segmentation in it and
eventually ends up having the final produced segments
appeared to be blockish. This happens to other image cells
when this similar situation occurred. Notice that the roller
coaster track is not segmented out together with blue
background. This shows the normalised cuts algorithm is able to tackle fine details of the objects in the image and also
across the image cells individually. Fine tuning the number
of clustered segments per cell gives an indicator such that to
what extent that the segmentation has to be accurately
performed. Increasing the number of segments per cell does
not significantly increase the computation time if the number
of cells to be divided for the image is constant. Fig. 10 shows
the effect of changing the number of clustered segments per
cell on the segmentation process while retaining the number
of image cells in the image. As the number of segments per
cell increased, more detailed contents of the cell are
segmented out. For the segmentation done with 10 clustered
segments per cell, the letter at the tail section of the aircraft is
segmented out. However, the number of clustered segments
per cell is increased to the extent where some of the edges in
the image such as the shading and shadow of the same
objects are also segmented out and considers them as distinct
segments. Thus, selecting an optimal number of clustered
segments per cell should be based on the image complexity
to reduce the tendency of the image being segmented
undesirably. Otherwise, customising number of image cells
for a certain region of the image can be assigned when some
objects in particular region of the image need more detailed
segmentation on it.