基于区域协方差矩阵融合深度的显著性检测方法

Visual Saliency Estimation Based on Region Covariances Combined with Depth Information

针对复杂背景下显著性检测的问题,提出了一种基于区域协方差融合深度信息的显著性检测方法。首先,分别对原图和深度图提取颜色、方向、位置及深度特征组成协方差矩阵,然后对不同尺度下的图像进行区域协方差距离计算构成子显著图,最后将子显著图结合成主显著图。在此基础上,应用中心偏见和人眼敏感度因子对显著图进行增强,提高显著区域检测能力。实验结果表明,提出的方法能有效检测出显著目标,且在主观和客观评价的各项指标上均优于其它模型。

In order to detect salient of complex scenes,this paper proposes a saliency detection method based on region covariances assisted by depth information.Firstly,features including color,direction,position and depth are extracted to form a covariance matrix from local patch of the original image and the corresponding depth map.Then,the sub-saliency maps are obtained by calculating the regional covariance distance of different scales of the images and combined into the main saliency map by fusion method.