Computing Image Salience

In recent years a large number of methods have been developed for computing saliency maps from images. The majority of them involve analysing the colour distribution (e.g. looking for rare patches of colour). Also, many involve performing a segmentation of the image, and analyse the similarity of regions in terms of colour, etc.

In contrast we have developed an extremely simple method that is based on edges alone, and does not require the potentially unreliable stage of image segmentation. Edges are computed using a pyramid, and the distance transform is applied to the edge magnitudes at multiple thresholds to generate an estimate of edge density, which we have found to often be a reasonable indicator of salience.

source image

salience map

overlaid thresholded salience map

salience map with central prior

overlaid thresholded salience map

For some data sets, including a central prior can generate better results, although this is not always the case (see the example above).

More details are given in:

P.L. Rosin, "A simple method for detecting salient regions", Pattern Recognition, vol. 42, no. 11, pp. 2363-2371, 2009. Post-print|DOI: 10.1016/j.patcog.2009.04.021

You can download code to implement the original method described in the paper as well as an extended version.

return to Paul Rosin's homepage