To avoid processing such large matrix, it is suggested
that a high resolution image be divided into equal size of
box-shaped image cells [11]. Since natural scene image has
irregular pixels distribution such that the frequency of image
pixels for particular range occurred unevenly throughout the
whole area of the image, a preliminary check is run on every
image cell to determine whether the particular cell requires
normalised cuts algorithm to perform on it. The first stage of
segmentation is begun by providing the k1 number of
segments, normalized cuts algorithm is then performed to
segment out k1 number of clusters for the particular cell. In
this segmentation stage, over-segmentation would likely to
be occurred as discrimination power in an image cell is
reduced compared with the discrimination power on the
whole image. Nevertheless, it helps to reduce the tendency of
important object boundaries to be missed out. In every image
cell, the implementation of simple k-means clustering plays
an important part on clustering out the generated eigenvector
is based on the given number of segments, k1.
Fig. 4 shows an image is divided into designated number
of cells and local segmentation on each of the cells. The
segmentation on each of the cells is done independently. In
other words, the segmentation done in a cell is not co-related
with the segmentation done in other cell. Segmented clusters
from the image cells are then used for second stage
segmentation.