Local structure in images from entropy production

Employing a thermodynamical model applied to scale-space transformations, we derive a method to encode different types of image features, as regions of different information content. Information is locally defined in terms of activity of the density of entropy production along scales. An activity map is formed, which represents the conspicuity or saliency of each point of the image. According to the different levels of activity, labelled as high, medium or low, we can extract basic parts of the image by collecting points of similar saliency.