Learn local priors by transferring training masks for salient object detection
Dan Wang, Canxiang Yan, Quan Zhou · 2017
In this paper, we present a novel framework to incorporate high-level guidance and low-level features to automatically identify salient objects based on two ideas. The first one considers the specific location prior to encode visual saliency, while the second one estimates image saliency using contrast with respect to background regions. The proposed framework consists of the following three steps: a) Top-down process: a specific location saliency map (SlSm) is learned. Specifically, for each image window (patch), a set of image windows with similar appearances are searched from a training image set and the corresponding segmentation masks are linearly integrated to produce a prior map, which guides the saliency object detection. 2) Bottom-up process: a multi-layer segmentation framework is employed, which provides vast robust background candidate regions specified by SLSM. Then the background contrast saliency map (BCSM) is computed based on low level image stimuli features. SLSM and BCSM are finally integrated to produce a pixel-based saliency map. Extensive experiments show that our approach achieves competitive results over MSRA 1000 and SED datasets, where each image contains more than one salient object.