RGB-D Image Saliency Detection via Background and Foreground Fusion

Abstract

RGB-D image saliency detection refers to the addition of depth information in traditional 2D images to extract significant objects. However, for current saliency detection models, most of them focus on the saliency objects themselves, but ignore the background content. Therefore, this paper proposes a novel saliency detection model that takes depth information into consideration of both background and foreground to extract salient area. Firstly, the foreground noise is removed by the background measurement mechanism of the image boundary information and the background seed is selected from the boundary superpixels to calculate a background-based saliency map. Secondly, the input image is constructed into a graph, and the depth information is introduced into the graph. The foreground seeds are obtained by using clues such as color, depth and position, and the foreground-based saliency map is calculated. Finally, the background and foreground map are merged to obtain the initial saliency map, the cell optimization is performed, and the iterative propagation is performed to obtain the final saliency map. The comparison experiments are carried out on three RGB-D image saliency detection datasets LFSD, NJU-400 and NJU-2000. The experimental results show that the proposed method is effective and improves the accuracy.