Simple Method for Detecting Visual Saliencies Based on Dictionary Learning and Sparse Coding

The human brain and the human vision system, in front of a given scene, focuses on regions with more information. Many pieces of research in the field of psychology, neuropsychology and cognitive neurosciences, have detailed the way in which the extraction of such information is carried out, even proposing models of the functioning of visual attention. From these models, computer scientists have proposed computational variants, which imitate human visual attention. This field of research is known as saliency detection on images. In this paper, a new and simple approach for detecting visual saliencies based on Dictionary Learning and Sparse Coding is presented. Our method first subdivides the image into full overlapped patches and runs a dictionary learning over them for obtaining its sparse representation. Then, by analyzing the sparse coding matrix, we compute how many image patches a dictionary atom affects in order to classify them as frequent or rare. Then, we calculate the saliency map of the image according to the composition of the image patches, i.e. an image patch is considered salient if it is mainly composed of rare atoms, an atom is rare whether it affects a few patches. Numerical results validate our method.