多角度融合的RGB-D显著检测

RGB-D saliency detection based on multiple perspectives fusion

显著检测是计算机视觉的重要组成部分,但大部分的显著检测工作着重于2D图像的分析,并不能很好地应用于RGB-D图片的显著检测。受互补的显著关系在2D图像检测中取得的优越效果的启发,并考虑RGB-D图像包含的深度特征,提出多角度融合的RGB-D显著检测方法。此方法主要包括三个部分,首先,构建颜色深度特征融合的图模型,为显著计算提供准确的相似度关系;其次,利用区域的紧密度进行全局和局部融合的显著计算,得到相对准确的初步显著图;最后,利用边界连接权重和流形排序进行背景和前景融合的显著优化,得到均匀平滑的最终显著图。在RGBD1000数据集上的实验对比显示,所提出的方法超越了当前流行的方法,表明多个角度互补关系的融合能够有效提高显著检测的准确率。

Saliency detection is an important part of computer vision and many creative studies have been proposed.However,most existing methods focus on detecting salient objects in 2D images and they cannot be used to detect salient objects in RGB-D images.In the paper,a new RGB-D saliency detection method based on multiple perspective fusion is proposed.The method considers multiple complementary relationships including color and depth,global and local as well as foreground and background to detect salient objects.Firstly,it combines color feature and depth feature to construct the graph model.Secondly,it computes the global and local compactness feature based on the graph model to get the initial saliency map.Thirdly,it uses the boundary connectivity based on color and depth to calculate the background probability as the weight on initial saliency map and then produces the background optimization saliency map.Finally,it extracts the foreground inquiry node from background optimization saliency map and then uses the manifold ranking algorithm to get the final saliency map.Experiments show that the fusion policy of multiple perspectives can effectively improve the accuracy of saliency detection and outperforms other state-of-the-art algorithms in RGBD1000 benchmark.