Editorial board editor-in-chief

parts of 125 images each, with an interval between them, what 
allowed the subject to rest. The sequence of images was random 
for each person, as well as the place of the emergence of the 
stimulus. SOA and intervals between successive images were 
random as well. For experiment generation and results analysis, 
scripts developed in the Matlab environment were used, and 
to perform the experiment dMdX program [9] developed at the 
University of Arizona was employed. For each of the 20 subjects 
300 measurements of reaction time were obtained (50 trial and 
250 test images). 
Saliency measure calculation
The experimental data were analyzed by comparing the reac-
tion times of the subject and the saliency measure for a region 
where the cue appeared. Because of the proposed methodol
-
ogy (Chapter 3.2) several methods for calculating the saliency 
measure in each of the nine image regions were designed and 
implemented. The methods were divided into two groups: based 
on the simulated gaze ixations mechanism and based directly 
on the saliency map. In the irst case a simpliied version of the 
gaze ixations prediction method proposed in [1] was used. The 
method consisted of two steps: inding the global maximum on the 
saliency map, recognizing it as a ixation point and then setting 
the conspicuity values in the surround of this point to 0, which 
simulated the inhibition of return mechanism. In the described 
manner the irst 5 and 30 ixations locations were determined. 
Subsequently, to each of the 9 image regions the saliency mea-
sure was assigned. In the irst method (method 1) the maximal 
value was set for the region with the irst ixation and for other 

---

53
T
elema
tics
Effectiveness analysis of selected attention models
regions the saliency measure was set to 0. The second method 
(method 2) was based on generating 30 successive ixations 
and calculating the saliency measure for each image region as 
an weighted sum. The irst ixation was assigned a weight of 30
2
and the following of 29
2
, 28
2
etc. 
Additionally two methods of calculating the saliency measure, 
based directly on the saliency maps, were proposed. The irst 
method (method 3) used as the measure mean saliency in a re-
gion. It was a simple approach, but its main drawback was that it 
favored regions with large areas of medium level of saliency. That 
is not consistent with the theory of attention attracting mechanism, 
in which small areas with high conspicuity are favored. The 
second measure (method 4) was based on calculating mean 
saliency in the surround (a circle of radius 26 pixels) of the global 
maximum for each region.
In the irst step of the performed data analysis, saliency maps for 
all 300 were computed using each of the ive considered models. 
Maps generated by the Zhai & Shah, SUN ICA, SUN DOG and 
Bruce & Tsotos methods were processed using the morphological 
top-hat operation [10] with a 25x25 square structuring element, 
which allowed to emphasize local maxima. For the last method: 
Itti, Koch, Niebur, which generates maps signiicantly different 
from others (Fig. 1), this procedure was not justiied. Then the 
maps were normalized to range (0;1) and the described saliency 
measures were calculated.
Results and discussion