Abstract:

As background area is similar to foreground region in RGB-D co-saliency detection algorithm and is easy to be classified as saliency region, an RGB-D co-saliency detection algorithm is proposed to select more accuracy seeds and reduce misclassification rate. Original image, depth image, and initial saliency map are input, which is obtained from existing algorithm, and Depth Shape Prior（DSP） is used to combine depth image and initial saliency map adaptively to generate a better initial saliency map. Objectness estimation is applied to pick out superpixels, which has higher possibility to be saliency region. Co-saliency propagation algorithm and cellular automaton optimization are utilized to get a more accuracy saliency map. Experimental results on RGBD Cosal150 dataset demonstrate the effectiveness and superiority of the proposed algorithm.

Key words: RGB-D co-saliency detection, objectness estimation, co-saliency propagation

摘要：

为了解决在RGB-D协同显著检测算法中和前景区域相似的背景部分易被分类为显著区域的问题。提出了基于目标检索挑选出前景概率更高的显著种子，减少误分类率的RGB-D协同显著检测算法。输入原始图片、深度图，及现有算法得到的最初显著图，进行超像素分割，利用DSP（深度形状先验）算法优化初始显著图得到更佳初始显著图。使用目标检索挑选出显著值更高且更有可能是显著物体的超像素，使用协同显著判断准则求得显著值。协同传播算法加以元胞优化被利用来得到更加准确的显著图。在RGBD Cosal150数据集上的实验表明了该算法的有效性和杰出性，取得了较高的准确度。

关键词: RGB-D协同显著检测, 目标检索, 协同传播