Competitive Learning Techniques for Color Image Segmentation
Aaron Mavrinac
University of Windsor
mavrin1@uwindsor.ca
Jonathan Wu
University of Windsor
jwu@uwindsor.ca
Xiang Chen
University of Windsor
xchen@uwindsor.ca
Kemal Tepe
University of Windsor
ktepe@uwindsor.ca
Abstract
A method for color image segmentation using a competi-
tive learning clustering scheme is examined, and some basic
improvements are made. Two important aspects of the color
image segmentation problem, namely color space selection
and oversegmentation, are discussed in the context of the al-
gorithm, with comments about suitability and effectiveness
of choices for various applications. A variety of settings are
tested and compared to highlight performance.
1. Introduction
Color image segmentation is an important part of many
image processing and computer vision problems, includ-
ing recognition, image retrieval, and compression. It has
long been recognized that segmenting image pixels in a
three-dimensional color space is a much different problem
than segmenting grayscale pixels (where one can use rel-
atively simple thresholding techniques). Clustering tech-
niques based on the least-sum-of-squares criterion, such as
the K-means algorithm [1], have been shown to perform
well in classifying multivariate observations like color im-
age pixels.
Uchimaya and Arbib [6] propose using the competi-
tive learning clustering technique, in which the positions of
weight vector units in the feature space are updated when
they “win” (e.g., are closest in Euclidean distance) among
the other units given a random input vector. To avoid the
problem of a few units monopolizing the input space, they
modify the algorithm to start with a single unit and divide
units based on a win-count threshold until the prescribed
number of units has been generated.
Some slight modifications to the original algorithm in
[6] yield improvements in the theoretical and actual speed,
as discussed in Section 2, and are therefore included in all
subsequent tests and comparisons.
2. Basic Algorithm Improvements
For reference, the modified competitive learning algo-
rithm from [6] is summarized here.
The parameters required are the total number of itera-
tionsNmax, the weight vector splitting threshold θt, and the
learning rate α. Experimental results from [6] offer values
of Nmax and θt which result in good solutions:
θt = 400
√
n (1)
Nmax = (2n− 3)θt(n + 7) (2)
Note that the sufficiency condition for n units to be gener-
ated is Nmax ≥ (2n− 3)θt.
A single weight vector W0 is placed at ~µ0 (the global
center of mass of the input set X), and its wincount variable
and NR (the current number of iterations) are initialized to
zero. The competitive learning algorithm then proceeds as
follows:
1. Select a random input vector X from X .
2. Find weight vector Ww such that the squared Eu-
clidean distance ‖X−Ww‖2 is a minimum for all W
(select one randomly in case of a tie).
3. Update Ww by4Ww = α(X−Ww).
4. Increment the wincount of Ww. If greater than θt,
reset the wincount and generate a new weight vector
equal to Ww.
5. Increment NR. If NR = Nmax, stop. Otherwise, re-
peat from 1.
The rest of this section explains two slight modifications to
the original algorithm which yield improvements in speed.
2008 Congress on Image and Signal Processing
978-0-7695-3119-9/08 $25.00 © 2008 IEEE
DOI 10.1109/CISP.2008.376
644