Study Of Hierarchical Saliency Detection Based On Probabilistic Boundaries

With the popularization of digital products and the fast development of Internet, image has already become an important information carrier. How to extract interest information from images is a hot research area in image processing and computer vision. Inspired by biological vision attention mechanism, saliency detection aims at locating the regions of interest(ROI) or salient regions in images. Saliency detection outputs a so-called “saliency map” image, where the intensity of each pixel indicates the probability of that pixel belonging to salient objects. Provided with saliency maps, we can allocate more computation resources to processing the salient regions.In this paper, we proposed a hierarchical saliency detection method based on probabilistic boundaries. Different from previous region-based saliency detection methods, we designed a hierarchical segmentation method via the probabilities of boundaries. Firstly, with an adaptive threshold, we eliminated the boundaries with probabilities less than that threshold. Next, we clustered the remaining boundaries into n groups by K-Means and selected the minimum probability of each group as a segmentation threshold. With different thresholds, n layers of region segmentation results varying from coarse segmentation to fine segmentation are generated. For each layer we extract, we computed its saliency map based on the proposed robust center bias, border bias, color dissimilarity and spatial coherence measures. After a linear weighted combination of multi-layer saliency maps and Bayesian enhancement procedure, the final saliency map is obtained.For a fair evaluation of our algorithm, we selected three challenging image datasets and several common evaluation criteria. We compared our algorithm with 8 saliency detection methods proposed in recent years. Numerous experimental results demonstrate that our proposed model outperforms 8 state-of-the-art saliency detection models. Meanwhile, with the obtained saliency maps, we perform object segmentation by Saliency Cut algorithm.