An entropy based ideal observer model for visual saliency

In this paper we present an information theoretic based model for saliency founded in the statistics of natural scenes. Our model uses the statistical models of natural images to estimate entropy as a measure of salience in those images. Theoretical models of the statistics of natural scenes have been used to solve a number of problems in image analysis and we demonstrate how it can be used to model fixation. We evaluate our model against human fixation data collected from a set of 100 images showing the value of incorporating image models based on the statistics of Natural Scenes for calculating salience. From our evaluation we find that our model identifies more locations in an image as being salient compared to other models, but for the most salient locations, which are fixated on by human subjects our model assign a higher salience value than other competing fixation models.