![](bg5.png)
In our experiments the result of searching for the
similar images is considered correct if within the
five most similar images either majority of them
belong to the same category. Otherwise an answer is
checked by an operator and is accepted if the
responded image visually bears a similar distribution
of dominating colors of the reference one. Although
the first criterion is easily measurable, the second is
subjective. However, the tests were performed by
three persons independently. The obtained results
are presented in Table 2.
The results show comparatively good
performance of the method since the achieved
accuracy is in the range of 82-89%. Usually, worse
results were obtained for more complicated scenes.
On the other hand, if an image consisted of few
objects with dominating color (such as in the
categories cars and faces), in majority of cases the
method was capable of selecting visually similar
instances.
Table 2: Results for image classification to different
categories.
Image category Accuracy
Cars 89 %
Flowers 87 %
Office 82 %
Faces 85 %
Very useful feature of the method is that it is
invariant to geometric deformations, as well as to
slight variations of illumination. This was measured
by artificially generated affinely transformed
versions of the reference images, for which
deformation parameters were randomly selected
from the predefined range. These were random
rotations of maximally 25, horizontal and vertical
changes of scale 12%, as well as translation of 25
pixels. To such deformed image additive noise was
added in the range of 10%. The algorithms for
generation of these deformations are described in the
book (Cyganek, 2009). The obtained results of these
tests show accuracy of 98-100%. The invariance to
the geometric deformations is mostly due to
measuring boundaries of dominating color
distributions, while to the variations of illumination
comes from the generalizing properties of the OC-
SVM classifiers.
4 CONCLUSIONS
The paper presents a simple but capable method of
the prototype encoding in a form of a set of
ensembles of OC-SVM classifiers. Such an encoding
allows fast examination of a database and selection
of images similar in their color distributions.
However, thanks to the boundary descriptors of the
OC-SVMs the output ensembles consume much less
memory than the original images or 3D histograms.
They also allow fast comparison of the test pixels
coming from the other images. The experimental
results show acceptable accuracy for three different
groups of test images.
Further research will be devoted to development
of methods that consider other characteristic features
of the images such as spatial position of color pixels
and texture. As alluded to previously, the presented
method should be connected with one of the search
methods that utilize invariant features of the images.
Future research should be also focused on
development of methods which allow responses
which agree with similarity in the sense of human
visual perception, as well as on human-computer
interfaces which allow easy formulations of queries
for search of visual information. For the latter, the
combination of different approaches seems to be the
most versatile, due to numerous categories of scenes
in the repositories. Also important is development of
parallel algorithms which allow faster operation for
very large databases.
ACKNOWLEDGEMENTS
This research was supported from the Polish funds
for scientific research in the year 2011 under the
Synat project.
REFERENCES
Aherne, F. J., Thacker, N. A., Rockett, P. I., 1998. The
Bhattacharyya Metric as an Absolute Similarity
Measure for Frequency Coded Data. Kybernetika, Vol.
34, No. 4, pp. 363-368.
Bertsekas, D. P., 1996. Constraint Optimization and
Lagrange Multiplier Methods.
Athena Scientific.
Bhattacharyya, A., 1943. On a Measure of Divergence
Between Two Statistical Populations Defined by their
Probability Distributions. Bull.
Calcutta Mathematic
Society
, Vol. 35, pp. 99-110.
Cyganek, B., Siebert, J. P., 2009. An Introduction to 3D
Computer Vision Techniques and Algorithms, Wiley.
IMAGE CONTENTS ANNOTATIONS WITH THE ENSEMBLE OF ONE-CLASS SUPPORT VECTOR MACHINES
281