Modeling Attention and Perceptual Grouping to Salient Objects

In this paper, we propose a biologically inspired model of
the middle stages of attention, with specific algorithmic details. Exist-
ing computational models of attention concentrate on their role in visual
feature extraction and the selection of spatial regions. However, these
methods ignore the role of attention in other stages. Extension of these
models has been proposed by augmenting the unit of attentional selection
to "proto-objects". In our approach, we extend attention to the middle
stages and integrate the selection process with the perceptual grouping
process. Integration is achieved by our innovative saliency driven per-
ceptual grouping strategy, extending the traditional pixel-based saliency
map to salient proto-objects. The proposed selective attention is made
in two stages. Firstly, to achieve salient region localization, our method
enhances the saliency map with region information from image segmen-
tation and selects the most salient region (proto-object). Then, regions
are organized using perceptual groupings, and their pop-out sequence is
determined. Compared with traditional attention models our model pro-
vides saliency maps with meaningful region information, by eliminating
misleading high-contrast edges, and focus of attention shifts in unit of
perceptual object rather than spatial region. These two improvements ¯t
to high stage vision information processing such as object recognition.
Experiments in a reduced set of images show that our proposed model
is able to automatically detect meaningful proto-objects.

ISBN:978-3-642-00581-7

Publication year:2009

Keywords:Visual attention mechanism, Perceptual Grouping, Salient object of interest