Hierarchical learning for salient object detection

Abstract

Most existing methods for salient object detection either depend on simple feature, such as contrast or boundary prior, which is sensitive to background variety, or extract redundant features for robustness, which is time-consuming. In this paper, we propose a hierarchical learning structure to alleviate the demanding feature. The hierarchical learning is based on low-rank (LR) decomposition and broad learning system (BLS). LR model with Laplacian constraint is applied to roughly separate foreground from background, which produces several positive and negative super-pixels as example to train a BLS classifier. The classifier is used to determine the final saliency of each superpixel. Experiments on two public datasets including MSRA10K and ECSSD show that our method achieves state-of-the-art result compared with the other nine methods.