Journal of Software Engineering and Applications, 2012, 5, 172-174
doi:10.4236/jsea.2012.512b033 Published Online December 2012 (http://www.SciRP.org/journal/jsea)
Biology Inspired Image Segmentation using Methods of
Artificial Intelligence
Radim Burget, Vaclav Uher, Jan Masek
Departmentof Telecommunication, Brno University of Technology,Brno, Czech Republic.
Received 2012
ABSTRACT
In recent years,many efforts have been devoted to image segmentation. Although for a man general image segmentation
is considered an easy task, for computers it is still considered to be difficult, computationally intensive and still unre-
solved task. This work presents an innovative algorithm combining theory of artificial intelligence and knowledge of
human eye anatomy. The resulting algorithm has not ambitions to be universal like human brain but can be trained and
perform on selected domain. The effectiveness of the algorithm is demonstrated on the selected examples.
Keywords: Image Processing; Image Mining; Feature Extraction; Data Mining
1. Introduction
In recent years, many efforts have been devoted to image
segmentation [1-3]. Although for a human the segmenta-
tion is relatively easy task, for computer it is still com-
plex, difficult, and computationally intensive and it is
still considered to be unresolved task.
Recently, there have been several efforts, which were
trying to create general segmentation algorithmthat is
trying to perform like a human [4,5]. These approaches
are computationally very intensive and also the accuracy
is not very high for most of the cases.
In many applications of the segmentation algorithms –
e.g. in the production line cases - there is not necessary to
perform well on general tasks. Specialized algorithms
can offer better accuracy and often lower computation
demands.
The main contribution of this paper is a novel image
segmentation method, which combines current state-of-
he-art artificial intelligence algorithms with a finding in
the field of biology of the human eye. The algorithm is
demonstrated on a selected task and compared with
thresholding and statistical region merging segmentation
[6]. The resulting software[11-3] was published online
and provided under open source license.1
The rest of the paper is structured as follows. Section 2
describes in more detail the problem being solved in this
paper, section 3 describes proposed segmentation model,
section 4 describes feature extraction from image data,
section 6 demonstrates performance of the model on se-
lected topics and section 7 concludes the paper.
2. Problem Definition
As indicated earlier in the text, although the segmenta-
tion is an easy task for a human, for computers it is quite
difficult. In Figure 1 a correct segmentation by a human
and incorrect segmentation performed by a statistical
region merging [6] is depicted. When a man processes
image, many aspects are taken into consideration includ-
ing the known and expected shape of the object.
In this paper a segmentation method based only on
pixel information and some feature extraction methods
will be used.
3. Segmentation Model
The overall principle of the model is depicted in Figure
2. It can be divided into two parts: a training part and a
performing part.
An objective of the training part is to prepare a trained
model, which can be then usedfor classification in the
performing part in future. When the training part is com-
pared to the performing part it is often relatively compu-
tationally intensive.It can take from a few seconds to
many hours of training.
An input of the training part is some training data.
These data are often called training data. On these data
usually a domain expert (e.g. in case of medical or some
specific domain data) has to classify the parts of training
data which are positive and which are negative. The out-
come of this process is so called labelled data. In order
1http://splab.cz/ts

Biology Inspired Image Segmentation using Methods of Artificial Intelligence
173
to obtain more information about the image pixels, some
feature extraction methods have to be performed on the
training images. The methods used for feature extraction
are described in detail in the section 4.
When the labelled training data and extracted features
are available, the image data can be transformed into
training data in a structured form. The structured form
data are considered a table data where each row stands
for each training sample (i.e. training point in the training
image) and each column stands for different feature ex-
tracted from the image.
When the training data are ready, the learning algo-
rithm can be used. For computer vision one of the most
interesting algorithms are: Decision Tree [7], Support
Vector Machines [8], Random Forest [9] and k-Nearest
Neighbours [10].
Figure 1.Segmentation of an image by human and computer.
Figure 2. Segmentation model principle.
4. Feature Extraction
In order to retrieve more information about the image
and their pixels so-called feature extraction was applied.
The feature extraction stands for application of different
transforms to an original image. The transform used were
Gaussian blur with different kernel sizes: σ = 1, 2, 4, 8,
16, 32:

22
2
2
2
1
,,
2
xy
Gxy e





Hessian curvatures with different scales σ = 2, 4, 8,
edge detection using Canny, Sobel and Prewitt filters.
The feature extraction about each pixel was motivated
by findings known from anatomy of human eye, which is
depicted in Figure 3. As seen from the figure, the density
of light receptors in the human eye has the highest den-
sity in the centre of the eye. Further from the centre of
the eye the receptors are denser and the picture there is
more and more blurred.
5. Model Optimization
In order to achieve better accuracy, forward selection,
backward selection was used in order. The forward se-
lection adds attributes to the model one at a time. At each
step, each attribute that is not already in the model is
evaluated for inclusion in the model. The attribute with
the best impact on the evaluation metrics is added to the
model. This is repeated until adding of new attributes has
no any improvement effect.
Backward selection starts with all candidate attributes,
tests the deletion of each attribute using a chosen model
comparison criterion and deletes the attribute (if any) that
improves the model the most. This is repeated until no
further improvement is achieved.
The idea behind cross-validation is to divide the data
sample into a number of k folds. Most often the number
of k equals 10 or equals the number of data samples (in
this case it is called leave-one-out cross-validation). We
train a learning model on k – 1 folds and the trained
model is appliedand evaluated on the fold, which were
missing during the training. This is repeated k-times so
the each fold is being left and evaluated. The resulting
evaluation is summarized results of the kevaluations.
Figure 3.Principle of receptors in human eye anatomy.

Biology Inspired Image Segmentation using Methods of Artificial Intelligence
174
The evaluation metrics can be arbitrary. In case of
image segmentation most often metrics are accuracy:
TP TN
TP TN FP FN
,
nn
Annnn


precision:
TP
TP TN FP FN
,
n
Pnnnn 
or root mean square error
02
TP TN FP FN
()
.
in
i i
yy
Ennnn



TP
n stands for number of true positively classified
pixels, TN
n stands for number of true negatively classi-
fied pixels, FP
n stands for number of false positively-
classified pixels, FN
n stands for number of false posi-
tively classified pixels, n stands for number of all classi-
fied pixels, ˆi
y stands for true value (-1 or +1) and
stands for classification and the confidence about the
classification (real number in interval -1 to +1).
6. Experiments and Evaluation
The evaluation examples of the proposed algorithm are
depicted in the figure 4. In the first example there is de-
picted segmentation of data taken from Google Earth
application in order to automate a measurement of build-
up area. The measurement takes about 30 seconds for an
image of dimension 1128x661 pixels.
The second example is biomedical data taken from
high resolution microscope depicting brain structure of
Drosophila[13]. The task here is to segment membrane
from the rest of the images in order to be able to analyse
micro- and macro- electrical circuits in the brain. Since a
single cubic millimetre contains for about 50 thousands
connection, for a human it is very time consuming task.
There is also another successful work that uses the
presented algorithms such as [14,15].
7. Conclusion
In this paper an innovative work dealing with trainable
image segmentation was presented. The work uses dif-
ferent feature extraction, machine learning and optimiza-
tion methods. The work [11-13|] is able produce high
accuracy segmentation models for variety of specific
tasks and was released as an open-source model2.
8. Acknowledgements
This work was supported by projects FR-TI4/151 and
VUT BD18116001, Czech Republic.
REFERENCES
[1] J. Tang, A color image segmentation algorithm based on
region growing, 2nd International Conference on Com-
puter Engineering and Technology (ICCET), 2010
[2] ChuanLong Li; Ying Li; XueRui Wu, Novel Fuzzy
C-Means Segmentation Algorithm for Image with the
Spatial Neighborhoods, 2nd International Conference on
Remote Sensing, Environment and Transportation Engi-
neering (RSETE), 2012
[3] Hasanzadeh, M.; Kasaei, S.; Mohseni, H., A New Fuzzy
Connectedness Relation for Image Segmentation, 3rd In-
ternational Conference on Information and Communica-
tion Technologies: From Theory to Applications, 2008.
ICTTA 2008.
[4] M Rousson, C Xu, A general framework for image seg-
mentation using ordered spatial dependency, Medical
Image Computing and Computer-Assisted Intervention –
MICCAI 2006, Lecture Notes in Computer Science, 2006,
Volume 4191/2006
[5] Yong Xia, A General Image Segmentation Model and its
Application, Conference on Image and Graphics, 2009.
ICIG '09. Fifth International
[6] R. Nock, F. Nielsen (2004), "Statistical Region Merging",
IEEE Trans. Pattern Anal. Mach. Intell. 26 (11):
1452-1458
[7] J. R. Quinlan, Induction of Decision Trees, Journal Ma-
chine Learning, 1986
[8] Cortes, Corinna; and Vapnik, Vladimir N.; "Sup-
port-Vector Networks", Machine Learning, 20, 1995.
[9] L. Breiman: Random Forests. Machine Learning 45(1):
5-32 (2001)
[10] Cover TM, Hart PE, "Nearest neighbor pattern classifica-
tion". IEEE Transactions on Information Theory 13 (1):
21–27.
[11] RadimBurget, PetrCika, Martin Zukal, Jan Masek: Auto-
mated localization of Temporomandibular Joint Disc in
MRI images. Telecommunications and Signal Processing
(TSP), 2011 34th International Conference on: p.: 413-41
[12] Burget, R., Uher, V., Masek, J., Trainable Segmentation
Based on Local-level and Segment-level Feature Extrac-
tion, 2012 IEEE International Symposium on Biomedi-
cal Imaging: From nano to Macro, 2012
[13] Uher, V., Burget, R. , Automatic 3D Segmentation of
Human Brain Images Using Data-mining Techniques,
35th International Conference on Telecommunications
and Signal Processing (TSP 2012), 2012
[14] Burget, R., Karasek, J., Smekal, Z., et al., RapidMiner
Image Processing Extension: A Platform for Collabora-
tive Research, 33rd International Conference on Tele-
communication and Signal Processing Location: Vienna,
AUSTRIA Date: AUG 17-20, 2010
[15] Burget, R.; Cika, P.; Zukal, M.,Automated Localization
of Temporomandibular Joint Disc in MRI Images 34th
International Conference on Telecommunications and
Signal Processing (TSP), Budapest, HUNGARY, AUG
18-20, 2011
2http://splab.cz/ts