基于带汇点Laplace扩散模型的显著目标检测

该文基于Laplace相似度量的构造方法，针对两阶段显著目标检测中显著种子的不同类型(稀疏或稠密)，提出了相应的显著性扩散模型，从而实现了基于扩散的两阶段互补的显著目标检测。尤其是第2阶段扩散模型中汇点的融入，一方面更好地抑制了显著性图中的背景，同时对于控制因子的取值更加稳健。实验结果表明，当显著种子确定时，不同的扩散模型会导致显著性扩散程度的差异。基于带汇点Laplace的两阶段互补的扩散模型较其他扩散模型更有效、更稳健。同时，从多项评价指标分析，该算法与目前流行的5种显著目标检测算法相比，具有较大优势。这表明此种用于图像检索或分类的Laplace相似度量的构造方法在显著目标检测中也是适用的。

Salient Object Detection Based on Laplace Diffusion Models with Sink Points

Based on Laplace similarity metrics, corresponding diffusion-based saliency models are proposed according to different clusters (sparse or dense) of salient seeds in the two-stage detection, a diffusion-based two-stage complementary method for salient object detection is therefore investigated. Especially for the introduction of sink points in the second stage, saliency maps obtained by this proposed method can well restrain background parts, as well as become more robust with the change of control factor. Experiments show that different diffusion models will cause diversities of saliency diffusion degree when salient seeds are determined. In addition, the two-stage Laplace-based diffusion model with sink points is more effective and robust than other two-stage diffusion models. Meanwhile, the proposed algorithm is superior over the existing five state-of-the-art methods in terms of different metrics. This exactly shows that the similarity metrics method applied to image retrieval and classification is also available for salient objects detection.