Bottom-up saliency model generation using superpixels

Prediction of human visual attention is more and more frequently applicable in computer graphics, image processing, human-computer interaction and computer vision. Human attention is influenced by various bottom-up stimuli such as colour, intensity and orientation as well as top-down stimuli related to our memory. Saliency models implement bottom-up factors of visual attention and represent the conspicuousness of a given environment using a saliency map. In general, visual attention processing consists of identification of individual features and their subsequent combination to perceive whole objects. Standard hierarchical saliency methods do not respect the shape of objects and model the saliency as the pixel-by-pixel difference between the centre and its surround. The aim of our work is to improve the saliency prediction using a superpixel-based approach whose regions should correspond to objects borders. In this paper we propose a novel saliency method that combines a hierarchical processing of visual features and a superpixel-based segmentation. The proposed method is compared with existing saliency models and evaluated on a publicly available dataset.