IMAGE SEGMENTATION USING GRAPH-THEORETIC CLUSTERING

摘要

In an object-based image modeling approach, scenes depicted in an image have to be segmented into meaningful objects or regions that share some similar features. Biological vision phenomena is ruled by principles such as perceptual grouping, selection, and discrimination, which mostly depend on regularities of nature such as colors, cohesiveness of matter or existence of bounding surfaces. Based on this information, we create a process that utilizes low level pixel processing and combines them with high level grouping to segment the image. We propose a method of image segmentation using near-clique windowing technique and graph-theoretic clustering method. This method processes an image at low level pixel values and groups these results at a global level. The final segmentation is not limited to one local scale, but may include information from a number of local scales. Successful application of this method to segment real world images is presented.