Region clustering with high level semantics for image segmentation

The task of image segmentation is to partition an image into disjoint and salient regions, which form meaningful objects. Traditional approaches mostly rely on similarities of low-level cues which can only identify objects with similar visual features. To get the entire complex objects, higher level information, such as co-occurrences of visual features and spatial information, is needed to overcome the semantic gap problem. However, the most attempts to integrate such semantics only use one of them or simple spatial relationships of image regions. In this paper, a region clustering method is proposed for image segmentation, in which co-occurring relationships are captured with LDA. And the quantitative spatial distances are incorporated in similarity graph construction for spectral clustering. Experiments showed that our algorithm with the introduced semantics achieved very good results on different kinds of images.