点击次数：

论文类型：期刊论文

发表时间：2017-11-01

发表刊物：VISUAL COMPUTER

收录刊物：SCIE、Scopus、EI

卷号：33

期号：11

页面范围：1467-1482

ISSN号：0178-2789

关键字：Nonlocal L-0; Regression predictor; Saliency detection; Interactive segmentation

摘要：Estimating salient object regions automatically has enhanced many computer vision applications in recent years. By observing the intrinsic sparsity of saliency map, we propose a graph-based nonlocal (NL) minimization framework to extract its sparse geometric structure. Our experimental results demonstrate that our method with artificially designed control map yields a significant improvement compared with the state-of-the-art saliency detection methods on four publicly available data sets. These saliency maps are further used for content-aware image resizing and unsupervised matting to test their uniformity. Moreover, we propose to learn the control map adaptively from training data. This strategy totally differs from the previously designed one, which is verified to be effective on image-classified data set. NL with data-driven strategy is extended to interactive segmentation task and is affirmed to be better-performed than other advanced interactive approaches.