Editorial board editor-in-chief

EFFECTIVENESS ANAlYSIS OF SElECTED ATTENTION mODElS
M
ałgorzata
w
łodarczyk
1
, t
oMAsz
k
ryjAk
2
, P
iotr
w
olski
3
1
 Multidisciplinary School of Engineering in Biomedicine, AGH-UST Kraków, Poland 
2
 Department of Automatics, Laboratory of Biocybernetics, AGH-UST Kraków, Poland 
3
Institute of Psychology, Jagiellonian University, Kraków, Poland
Abstract:
The aim of this study was an empirical veriication of ive different computer models of exogenous human attention. 
These models allow to predict the likely ixations on a static image and calculate saliency maps which show how distinctive the 
image elements are.
The basis for the veriication of the models was a modiied cuing task, originally developed by Posner. Using the 
algorithms provided by the authors of the models an analysis of selected images was performed. Then the effective
-
ness of the models was veriied in experimental studies: an image was presented to the subject, then on this image 
a visual stimulus appeared and the reaction time was measured. It was assumed that the emergence of a stimulus 
at the point where previously an attention attracting element appeared would accelerate the reaction.
The study was conducted on a set of over 300 different, emotionally neutral, natural images. Obtained results indi
-
cate the potential usefulness of the proposed schema for testing the eficiency and reinement of attention models. 
Keywords:
attention modeling, saliency, the Posner’s paradigm 
item attracts attention on the basis of a “pop-out” mechanism: it 
somehow pops up, stands out from the background. Probably this 
is computed in terms of centre-surround mechanisms, which are 
a neuronal response to image differences between a center and 
a broader surround. Moving eyes for a particular, salient element 
is fully automatic and takes about 20-50 ms [2].
In recent years, studies and researches of mathematical, 
computational models of visual attention are developing [5]. Most 
researchers focus on the analysis and simulation of bottom-up 
attention processes. Much less work was done in the ield of 
top-down attention [2]. Models of bottom-up attention, consisting 
on an analysis of the image characteristics without taking into 
account the knowledge of the observer, are easier to construct. 
Numerous, detailed mathematical models that allow to compute 
so-called saliency maps, depicting image elements standing 
out between others, thus in the assumption attention-grabbing, 
have been already established. The effectiveness of the models 
is usually veriied in terms of free viewing, by comparing the hu
-
man ixations, obtained from the eye-tracking device, with the 
results of model calculations. It is assumed that the movement 
of the eyes follows attention. Some of the models concentrate 
on simulating brain processes, when the main aim of the others 
is to obtain results similar to those recorded from subjects [3]. 
Introduction
One of the most astonishing human capacities is the ability to 
analyze the visual ield in real time. Intermediate and higher 
visual processing allows to select the most relevant information 
from all available visual stimuli which are then further processed. 
Searching can be guided by the distinctive physical features 
of image elements (bottom-up model of attention), as well as 
can base on subject’s knowledge and depend on the task and 
desired item (top-down model of attention [1]). The models of 
top-down and bottom-up attention are two essential components 
in a contemporary frame-work for attention deployment [2]. 
Some researchers claim that the top-down attention, based 
on variable selection criteria, is probably more powerful than 
the bottom-up attention [2]. It is dependent on the task that is 
performed and on sought elements in the visual ield. What is 
more, it can be volitionally controlled and much of attention is 
guided by higher brain areas [2][11]. However, volitional control 
is connected with a timing cost and it takes 200 ms or more to 
move the eyes to the interesting element [2].
In contrast, the bottom-up attention is a response to a sheer, 
severe sensory stimulation [11]. Some elements in the visual 
ield stand out from the background automatically, regardless of 
a subject and an experience or a task [2]. The saliency of an ele
-
ment depends on the background and the context [12]. Particular 

---

T
elema
tics
50
Effectiveness analysis of selected attention models
Such models have many potential applications. They can be used 
in artiicial vision systems, robotics, navigation systems as well 
as in automatic detection of desired objects on the image or in 
space [2]. In addition, they may help in the detection of certain 
diseases and human disorders. 
In this study a speciic approach to attention models evalua
-
tion was proposed and tested. It relies on the Posner’s cuing task 
[4]. It allows testing attention models without the eyes movement 
registration. In the original Posner’s task examined subjects had 
to focus eyes on the ixation point and to react without eyes 
movements as quickly as possible when the light stimulus ap-
peared. Some test ran with a clue, which pointed to the expected 
locus of the stimulus. If the pointer showed the correct location, 
the response times was shorter. If the pointer gave the wrong 
information, then the longer reaction time was observed. In this 
study the intensity of saliency was treated as a clue. Thus the 
basis for the evaluation of models was an assumption that more 
accurate models allowed a better prediction of reaction time in 
regions of different levels of saliency. 
Models of attention
It this study 5 saliency models were analyzed and veriied: from 
simple, based on image color statistics, to more complex ones, 
which are based on various image characteristics and simulate 
learning mechanisms. The following chapter contains a short 
description of each model putting the main emphasis on the 
manner of action.