Image segmentation with hierarchical topic assignment

Image segmentation can be viewed as the problem of topic assignment, in which pixels are grouped into regions with respect to the topic model of textures and real-world semantic meanings. In this paper, we apply a novel hierarchical topic model, which builds a multi-level image representation covers texture and semantic region, to image segmentation. Specifically, the proposed method firstly assigns texture-topic to each pixel according to the low-level model defines on visual word and neighboring constraint. Furthermore, inspired by the connection between image patch and linguistic sentence, we model semantic segments of natural scene as the combinations of texture topics. Finally, the segmentation result is achieved by finding homogeneous regions in topic field. Experimental results on natural scene images demonstrate the effectiveness of our method.