An important problem which is considered in this thesis is that of exploiting a model of visual attention in order to extract features which are not domain dependent. While this thesis is not concerned with developing a model for biological visual attention per se, we look to human psychophysics in order to motivate our particular approach to robust and efficient feature extraction.
As we have previously noted, we extract landmarks on the basis of local maxima of edge density. Work on human visual attention suggests that a key attribute of the loci of attention is that they are different from their surrounding context [30, 51, 58]. Several featural dimensions have been identified that lead to pre-attentive ``pop-out'' and, presumably, serve to drive short-term attention [57]. Probable feature maps used by human attention may include those for colour, edge density, or edge orientation. Other research demonstrates that attentional processing is characterised by visual saccades to areas of high curvature, or sharp angles [45]. Work by Dudek and Bourque demonstrates that the behaviour of an edge-density attention operator on simple stimuli resembles that predicted by the psychophysical literature [12], and is the basis for the operator employed in this work.
The next chapter will expand further on the idea of employing statistical extrema for feature extraction. We will present a formal definition for our attention operator, and further motivate our approach over traditional approaches to feature extraction.