Multi-Features Fusion Based on Boolean Map for Video Saliency Detection

Abstract

Video saliency detection aims to extract salient objects in video. In order to resolve the problems of incomplete extraction salient area and the mixing between background and salient objects at boundary, this paper proposes a video saliency detection method via multi-features fusion based on Boolean Map. The main innovation in this method is that global topological relationship of video frames are utilized to compute saliency values. Firstly, boolean maps are generated according to the Boolean Map theory via combine motion features and color features in videos, and then attention maps are calculated by masking unsurroundedness area, finally, video saliency maps are obtained by fusing all attention maps. Experiments on SegtrackV2 and Fukuchi benchmark datesets show that the proposed method successfully obtains complete and clear boundary salient area, and it outperforms the other general models.