Incorporating local image structure in normalized cut based graph partitioning for grouping of pixels

Keywords: Perceptual grouping Early human vision Image pixel grouping Local image structure Graph partitioning Normalized cut a b s t r a c t Graph partitioning for grouping of image pixels has been explored a lot, with normalized cut based graph partitioning being one of the popular ones. In order to have a credible allegiance to the perceptual grouping taking place in early human vision, we propose and study in this paper the incorporation of local image structure/context in normalized cut based graph partitioning for grouping of image pixels. Similarity and proximity, which have been studied earlier for grouping of image pixels, are only two among many perceptual cues that act during grouping in early human vision. In addition to the said two cues, we study three other such cues, namely, common fate, common region and continuity, and find indications of local image structure utilization during grouping of image pixels. Appropriate incorporation of local image structure/context is achieved by representing it using neighborhood in the form of histogram and fuzzy set. We demonstrate both qualitatively and quantitatively through experimental results that the incorporation of local image structure improves performance of grouping of image pixels. Image segmentation and the allied issue of grouping of image pixels have received a lot of attention of researchers over the last half a century. Although both image segmentation and grouping of image pixels refer to appropriate partitioning of images, we, in this paper, would like to have a distinction between them based on the cues driving the processes. As mentioned in [37], we consider image segmentation as a process that partitions images based on lower-level cues such as coherence of brightness, color and texture, and spatial proximity along with higher-level cues such as shape, model and knowledge of objects/regions. On the other hand, we consider grouping of pixels in images as a process that partitions images based only on lower-level cues, that is, image segmentation based on lower-level cues alone. Therefore, we would expect image segmentation to partition images into meaningful regions and grouping of pixels to partition images into homogeneous and spatially compact regions. We consider grouping of image pixels in the spirit of perceptual grouping in early human vision. Perceptual grouping in early human vision refers to the human visual ability to group image pixels by inferring relations among them from lower-level primitive image features without any knowledge of the image content [22,26]. …