Unsupervised visual saliency detection via information content measuring

Based on the philosophy that exploits image information content as the metric of visual saliency, an innovative method for unsupervised visual saliency detection is proposed. In the foundation of clustering input into semantically consistent regions, Shannon entropy and normalised pseudo-Wigner-Ville distribution are utilised for the measuring of image information content. As a consequence, an information content map can be obtained, and it is taken as a saliency indicator. Dynamic scale analysis is performed to establish saliency maps which contain well-defined salient object boundaries and efficiently suppressed background. Experiments on various cluttered natural images demonstrate the effectiveness of the proposed method.