基于视觉显著性的非监督图像分割

Unsupervised Image Segmentation Based on Saliency Detection

交互式的图像分割算法需要用户输入先验信息,从而增加了算法的时间复杂度和用户的负担。提出了基于视觉显著性的非监督图像分割算法。该算法首先通过均值漂移算法先对图像进行预处理,将图像过分割成互不重叠的小区域。这些区域采用区域邻接图表示,当两个区域相邻时对应的节点之间存在边。其次,通过计算各个区域的颜色相异性和纹理一致性,得到相邻区域之间的合并概率。再次,根据区域的颜色和空间位置信息,定义每一个区域的显著性指标,选择最大显著性指标对应的区域作为目标种子区域,图像边缘区域中显著性指标最小的区域作为背景种子区域。最后,基于最大相似性合并策略,对与种子区域相邻的且合并概率最大的区域进行合并。实验表明,所提算法不需要先验信息,且可以得到较好的分割效果;与非监督图像分割算法相比,所提算法可以避免过分割。

Interactive image segmentation needs the user interactions which increases the time complexity and the user＇s burden. We proposed an unsupervised image segmentation algorithm based on visual saliency. First, mean shift （MS） algorithm is used to obtain initial segmentation without overlapping. The regions generated by MS are represented by a region adjacency graph （RAG） and an edge exists only if two regions are adjacent. Second, the color dissimilarity and texture consistency between the regions are computed, which are adjacent, as the weight of the edge in our RAG. Then, the proposed algorithm defines the saliency index （SI） according to the color and spatial information of each region gene- rated by MS algorithm. The region with maximal SI is defined as the seed of object, and the region with minimal SI in the boundary is defined as the seed of background. Finally, region merging is performed according to the strategy of maximize similarity around the seed of object and background. The results show that the proposed algorithm obtains better segment results without any interactive information and avoids oversegmentation compared with other unsupervised image segmentation.