Salient object segmentation based on depth-aware image layering

Abstract

This paper proposes an efficient salient object segmentation method via depth-aware image layering. First, based on the multiscale region segmentation results of an input color image, the depth consistency integration is utilized to generate the image pre-segmentation result. Then, under the guidance of the depth histogram division, the pre-segmented regions are divided into several different layers to differentiate salient object regions and background regions. Finally, an adaptive sample update and selection method based on layered image regions is used to select appropriate training samples for salient object segmentation. The depth information of the image is fully utilized in each step of the entire framework. Experimental results on two public datasets demonstrate that the proposed method achieves the better performance than the state-of-the-art depth-aware salient object segmentation methods.