Self-Organizing Maps and Their Applications in Image Processing, Information Organization, and Retrieval - Nonlinear Signal and Image Processing: Theory, Methods and Applications - page 437

inspecting Table 11.2, one can see that the proposed single-layer split-merge
SOM achieves a slightly better performance than the standard SOM that uses
the same initialization with LBG, but with one fundamental difference: The
split-merge SOM algorithm has found the number of clusters present in the
input training set, while the standard SOM has been initialized in an optimal
way for 256 output neurons. The two-layer split-merge SOM architecture is
expected to give the best result when there is strong overlap between the
training subsets. It seems that this is not the case in our experiment. If the two
single-layer split-merge SOMs in the first layer were trained in parallel, then
the two-layer split-merge SOM architecture would provide almost identical
results with a (single-layer) split-merge SOM, but in half the computation
time.

11.7.4.2 Image Segmentation
Segmentation algorithms are usually employed to generate indices for image
retrieval. For example, a hierarchical tree-structured SOM is used for indexing
in the PicSOM system. 74 The evaluation of image segmentation algorithms
is a difficult task, mainly due to the absence of a clearly defined criterion of
“success.” It also affects the evaluation of retrieval systems. In this section
we compare the performance of the proposed split-merge SOM algorithm
in image segmentation to that of the adaptive iterated conditional modes
(ICM) algorithm. 75 , 76 The split-merge SOM is applied to five-dimensional

T . The training data set
is obtained by sampling the entire image with a step of 10 pixels per row,
every 10th row. All statistical tests were performed at 95% significance level.
The adaptation step size

input vectors x ( i, j ) = ( x R ( i, j ) ,x G ( i, j ) ,x B ( i, j ) ,i, j )

α = . ϑ = .

01. The algorithm starts with one
cluster. The ICM algorithm is applied on the luminance component of the
image. The ICM algorithm starts with four regions.
The nonzero value of the Pott's potential function was set to 2.5. Six images
from the collection of paintings of the Bridgeman Art Library were used in
the study. 77

0 2 and 0

The ground truth for the test images is given as a quantitative

description:

Image 107931— five objects: sky, mountains, people, foreground, water
Image 122100— four objects: figure, reflection, hammerhead cloud, the
rest
Image 134975— five objects: sky, buildings, trees, people, pavement
Image 47794— three objects: sky, people, hat
Image 72398— four objects: sky, trees, waters, sails
Image 98640— three objects: horses-sledges-people, snow, sky

Using the aforementioned verbal ground truth, the number of the obtained
regions by both algorithms is presented in Table 11.3.