next up previous
Next: References Up: Inform. theory approach to Previous: 8 Implementation

9 Discussion and future work

We have presented a framework based on an information theoretic contrast for decomposing image information in both scale space and spatial coordinates. The central representation is a Kullback contrast infinitesimal indexed in three dimensional scale space. It has a variety of uses. Integrated over spatial coordinates it gives an information distribution in scale space, for instance useful for a scale selection problem. Integrated over the scale dimension it is a base for attention selection in spatial coordinates.

The attention selection process and the underlying saliency maps are inherently task dependent. The saliency map need reflect what our task deems important in the image, and the attention selection process need reflect the temporal and importance order in which an agent need information. Our approach abstracts out these task dependencies and the information theoretic contrast can be applied to any normalized pre-attentive saliency map, and subsequently any task dependent attention selection can be applied to the scale space contrast measure. The use of a contrast and line based pre-attentive saliency map, as well as the simple thresholding in the spatial contrast can be seen as an example only to motivate the information theoretic framework.

We plan to explore the use of our information measures and attention selection method, in various practical applications. In visual space robotics visual features need to be found, and attended to during manipulation tasks [17]. Similarly model free, filter based methods used for visual search initially have to be pointed to interesting regions. So far a human operator has had to point out the relevant features on objects to manipulate, or objects which should be found. We intend to try to do this in an autonomous fashion with our attention selection procedure.

For the attention selection process we plan to investigate better methods to select image regions than just thresholding in the information distributions.

Another important issue is to find different primitive measures, which have some useful interpretation in the real world, to base our probability distributions on. Computing the information distribution over probability distributions with unclear interpretation could of course also yield practically useful results, but then the information theoretic approach we have taken has little advantage over any heuristic method.

Acknowledgments

To Jim Müller, Randal Nelson, who also provided many of the test images, Chris Brown, and Choh Man Teng for discussions on various aspects of this project.


next up previous
Next: References Up: Inform. theory approach to Previous: 8 Implementation