抄録
We propose a method for segmenting a color image and extracting a region of an object in a scene by using information which is not specific to the target object but common to a projection of the object in a space. The proposed method is based on a split-and-merge technique. The image is initially segmented into small regions by clustering in color space. Then the regions are merged based on four features of information on the configuration, connection and area ratio of adjacent pairs of regions. This merging allows us to extract an object as a single region even if the surface of the object is textured with a variety of reflectances; this isn't considered by conventional segmentation methods. We give our algorithm for segmentation and extraction, demonstrate experimental results, and make comparisons with some other methods.
