Segmentation Using Spatial Context and Feature Space Cluster Labels.

摘要

The focus of this paper is on image segmentation processes, collectively referred to as a 'low-level' vision system. The programs which are discussed here transform a large spatial array of pixels (picture elements) into a more compact representation through the exploitation of visual features, e.g., intensity, color, texture, etc. The goal is to detect a relative feature invariance across an area of the image and then to label all the pixels in any such area as belonging to the same region. Regions can be detected through global analyses (e.g., histogram clustering) which find interesting areas by ignoring the local textural configurations of the data, in conjunction with local anlayses (e.g., relaxation) which act as a fine-tuning mechanisms both to resolve global ambiguities and to accurately delimit region boundaries. (Author)