Heuristic analysis influence of saliency in the color diversity of natural images

The estimation of chromatic diversity of natural images is commonly quantified through the computation of the number of discernible colors and has received much attention because of the different implications it has. However, the relationship between that number and the number of colors that really attracts the attention from an observer is still not clear and has been given little attention. New concepts about salient discernible colors-the salient chromatic diversity of images- and remarkable salient colors-connected colors in the same salient image area-are introduced as opposed to the classical number of discernible object colors, which is usually evaluated for the global image without differentiating between probable attended and non-attended image regions. We have used different well-known saliency models to locate the salient regions in the scenes and have heuristically studied the extent to which those models preserve the chromatic diversity of natural images. Based on a bottom-up approach, a reduction of around 40%-55% in the number of discernible colors were obtained, and not all saliency algorithms preserved a uniform sampling of the original color gamut. Thus, our results suggests that particularly the graph-based visual saliency model got good low dissimilarity values in comparison with other approaches that put emphasis solely on color as the main low-level feature. Furthermore, we have introduced a quantification scheme of the average number of remarkable salient colors appearing in the images, and have proved how the heuristic-based analysis of salient image areas can be used to create segmented images automatically according to their salient chromatic diversity.