基于低秩背景约束与多线索传播的图像显著性检测

Image Saliency Detection Based on Background Constraint of Low Rank and Multi-cue Propagation

针对当前基于流形排序的显著性检测算法缺乏子空间信息的挖掘和节点间传播不准确的问题,该文提出一种基于低秩背景约束与多线索传播的图像显著性检测算法。融合颜色、位置和边界连通度等初级视觉先验形成背景高级先验,约束图像特征矩阵的分解,强化低秩矩阵与稀疏矩阵的差异,充分描述子空间结构信息,从而有效地将前景与背景分离;引入稀疏感知和局部平滑等线索改进传播矩阵的构建,增强颜色特征出现概率低的节点的传播能力,加强局部区域内节点的关联性,准确凸显节点的属性,得到紧密且连续的显著区域。在3个基准数据集上的实验结果与图像检索领域的应用证明了该文算法的有效性和鲁棒性。

Considering the lack of subspace information digging and inaccurate propagation between nodes in existing saliency detection algorithm based on manifold ranking,an image saliency detection algorithm based on background constraint of low rank and multi-cue propagation is proposed.Primary visual priors such as color,location and boundary connectivity prior are fused to form a background high-level prior,which restrains the low rank decomposition of feature matrix and strengths the difference between low rank matrix and spares matrix,describes structural information of subspace fully to separate foreground and background efficiently.Cues of rareness perception and local smoothing are introduced for improving the reconstruction of propagation matrix,which improves the node’s propagation capacity that has low probability of color feature occurrence,enhances the relevance of local region,strengthens the properties of nodes accurately to obtain the compact and continuous salient regions.The experimental results on three benchmark datasets and the application to image retrieval demonstrate the efficiency and robustness of the proposed algorithm.