Editorial board editor-in-chief

Bruce & Tsotos [8]
The model is based on the premise that the highest saliency 
values are assigned to the most informative regions of the im-
age. Therefore, it is analyzed how the content of the given image 
region is unexpected regarding to its surroundings. The conspi
-
cuity map is calculated using the Shannon’s Self-Information: 
-log(p(x))
, where 
p(x)
is the overall likelihood. In this method, for 
the whole image, responses to several ICA ilters are determined 
and then an estimate of the probability distribution is created. As 
the most salient are recognized these image regions where the 
joint likelihood of ilter responses is low. Details of this method 
are described in [8]. It is worth to notice that this method is similar 
to the one described in Chapter 2.2. In both models, as the most 
salient are recognized regions in which the response to ilters 

---

51
T
elema
tics
Effectiveness analysis of selected attention models
differs much from the prior calculated probability distribution. 
In the SUN method the distribution is calculated on the basis 
of a training set of images and in the Bruce & Tsotos method 
on the basis of the analyzed image. Additionally, authors of the 
method noticed that the proposed computational scheme results 
in an architecture with cells and connectivity reminiscent of that 
appearing in the primate visual cortex.
Itti, Koch & Niebur [1]
The saliency model proposed by L. Itti, C. Koch and E. Niebur 
from California Institute of Technology in 1998 is one of the most 
frequently quoted and evaluated models for calculating conspicu-
ity maps of images. The basis of the method is an assumption that 
attention is directed in different parts of the visual ield according 
to the image parameters (intensity, color, orientation, texture). The 
input image is analyzed using three features maps: intensity, color 
and orientation. Each map is then scaled to nine spatial resolu
-
tions using dyadic Gaussian pyramids (from scale 1:1 to 1:256). 
The features are computed by a set of linear center-surround 
operators akin to visual receptive ields. Typical visual neurons are 
most sensitive in a small region (center), while stimuli presented 
in a broader, weaker antagonistic region (surround), concentric 
with the center, inhibit the neuron’s response. Center-surround 
is implemented in the model as the difference between ine and 
coarse scales. The center is a pixel at scale 
c
∊
{2,3,4}
, and the 
surround is the corresponding pixel at scale 
s = c+
d
, 
d
∊
{3,4}. 
The across-scale difference between two maps is obtained by 
interpolation to the iner scale and point-by-point subtraction.
Given the input image in RGB color space the following 
features are considered: intensity (
I = (r+g+b)/3
) and colors: 
red (
R = r -(g+b)/2
), green (
G = g -(r+b)/2
), blue (
B = b -(r+g)/2
) 
and yellow (
Y = (r+g)/2 – |r-g|/2 – b
). The irst set of feature 
maps is connected with intensity contrast, the second with 
color channels (so-called “color double-opponent” system) and 
the third with orientation (iltering the intensity map with Gabor 
ilters of different orientations). In total, 42 feature maps are 
computed: 6 for intensity, 12 for color and 24 for orientation.
The saliency maps of each feature are then combined 
using a special normalization operator. It globally promotes 
maps in which a small number of strong peaks of saliency 
is present while globally suppresses maps which contain 
numerous comparable peak responses. Finally, the algorithm 
returns three conspicuity maps (intensity, color, orientation) and 
a global saliency map. All details of the algorithm are described 
in [1]. Additionally, Itti et al. proposed a method for predicting 
successive gaze ixations using local maxima of the saliency 
map and a 2D layer of leaky integrate and ire neurons. 
Methods