Saliency Detection via Robust Seed Selection of Foreground and Background Priors

Abstract

Recently, saliency detection has become a research hotspot in both the computer-vision and image-processing fields. Among the diverse saliency-detection approaches, those based on the foreground and background-based model can achieve promising performance. Reliable seed selection for the foreground and background priors is a critical step for successful saliency detection. In this paper, we firstly exploit the spatial distribution of the extracted directional patches to predict the centroid of each salient object in an image. Then, we adopt the located centroids as the visual-attention center of the whole image to compute the superpixel-based center prior, which can facilitate the subsequent seed selection for the foreground and background-prior generation. Finally, the two individual foreground-based and background-based saliency maps are combined together into a plausible and authentic saliency map. Extensive experimental assessments on publicly available datasets demonstrate the effectiveness of our proposed model.