A semantics based segmentation algorithm for scene images

This paper proposes and evaluates a new semantics based segmentation algorithm for scene images. This algorithm has two phases: the initial segmentation based on color-spatial information and region merging based semantics information. The initial segmentation uses an automatic region growth method without the need of seeds. It avoids the loss of detailed color information of the scene images by using non-quantized colors in region growth. This paper also innovatively includes the underlying semantics of the scene images during the region merging. The experiments show the new segmentation algorithm is very efficient and effective and could get a very accurate segmentation for scene images. The main regions also match well to people's visual perception.