Biomedical Engineering Reference
In-Depth Information
Itti et al. [23, 24, 25] recently proposed a purely bottom-up model, in which spatial
competition for salience is directly modelled after non-classical surround modulation
effects. The model employs an iterative scheme with early termination. At each
iteration, a feature map receives additional inputs from the convolution of itself by
a large difference-of-Gaussians filter. The result is half-wave rectified, with a net
effect similar to a winner-take-all with limited inhibitory spread, which allows only
a sparse population of locations to remain active. After competition, all feature maps
are simply summed to yield the scalar saliency map. Because it includes a complete
biological front-end, this model has been widely applied to the analysis of natural
color scenes [25]. The non-linear interactions implemented in this model strongly
illustrate how, perceptually, whether a given stimulus is salient or not cannot be
decided without knowledge of the context within which the stimulus is presented.
Many other models have been proposed, which typically share some of the com-
ponents of the three models just described. In view of the affluence of models based
on a saliency map, it is important to note that postulating centralized control based
on such map is not the only computational alternative for the bottom-up guidance
of attention. In particular, Desimone and Duncan [15] argued that salience is not
explicitly represented by specific neurons, but instead is implicitly coded in a dis-
tributed modulatory manner across the various feature maps. Attentional selection is
then performed based on top-down weighting of the bottom-up feature maps that are
relevant to a target of interest. This top-down biasing (also used in Wolfe's Guided
Search model [59]) requires that a specific search task be performed for the model to
yield useful predictions.
19.4
Top-down modulation of early vision
The general architecture for the bottom-up control of attention presented above opens
two important questions on the nature of the attentional bottleneck. First, is it the
only means through which incoming visual information may reach higher levels of
processing? Second, does it only involve one-way processing of information from
the bottom-up, or is attention a two-way process, also feeding back from higher
centers to early processing stages?
19.4.1
Are we blind outside of the focus of attention?
Recent experiments have shown how fairly dramatic changes applied to a visual
scene being inspected may go unnoticed by human observers, unless those changes
occur at the location currently being attended to. These change blindness experi-
ments [39, 41] can take several forms, yielding essentially the same conclusions.
One implementation consists of alternatively flashing two versions of a same scene
separated by a blank screen, with the two versions differing very obviously at one
Search WWH ::
Custom Search