Abstract: Co-saliency detection refers to finding common salient foreground region in a set of related images. The existing method captures the relationship of node pairs in the image, and uses human prior knowledge to construct a co-saliency detection model. However, the high-order relationship between the nodes in the detection is neglected, and the potential connection between the nodes is not mined, resulting in a suboptimal co-saliency map. This paper proposes a novel hypergraph co-saliency propagation detection framework for salient seed propagation. Specifically, the framework consists of inter saliency propagation and intra saliency constraints. The former uses salient seed point based on single saliency map and cross-image cross-propagation mechanism. The fusion algorithm detects the co-salient objects of the image group and suppresses non-co-salient object. The object obtains the preliminary co-saliency map. The later introduces the convex hull prior mechanism of the image, learns the spatial distribution information, constrains the common background noise, suppresses the similar background area, and obtains a more accurate co-saliency map. A large number of experiments are carried out on two widely used co-saliency detection datasets. The results show that as the same unsupervised model, compared with the existing unsupervised co-saliency detection method, this paper obtains the optimal performance.

Key words: co-saliency detection, inter saliency propagation, intra saliency constraint, convex hull prior

摘要： 协同显著性检测是指在一组相关图像中发现共同的显著前景区域。现有方法捕获图像中节点对的关系，利用人类先验知识构建协同显著性检测模型，然而忽略了检测中节点之间的高阶关系，没有挖掘到节点间的潜在联系，从而导致次优的协同显著图。提出了一个新颖的基于超图的种子传播的协同显著性检测框架。具体来说，框架由交叉图像显著性传播和图像内显著性约束组成，前者利用基于单显著图的显著种子点，跨图像交叉传播机制，融合算法检测图像组的协同显著对象并抑制非协同显著对象，获得初步协同显著图；后者再引入图像的凸包先验机制，学习空间分布信息，约束共同背景噪声，抑制相似背景区域，获得更精确的协同显著图。在两个广泛使用的协同显著性检测数据集上进行大量的实验，结果表明，同为无监督模型，相比现存的无监督协同显著性方法，获得了最优的性能。

关键词: 协同显著性检测, 交叉图像显著性传播, 图像内显著性约束, 凸包先验机制