Region Extraction and Description through Planning.
Abstract
This paper examines several image segmentation algorithms which have been explored in the development of the VISIONS system. Each of these algorithms can be viewed as a variation on a basic theme: the clustering of activity in feature space via histogram analysis, mapping these clusters back onto the image, and then isolating regions by analysis of the spatial relationships of the cluster labels. It is shown that the interaction between these two representations of data (global feature information and spatial information) provides a view that is lacking in either. The scene segmentation algorithms contain the following stages: (1) PLAN: reduce the amount of detail in the scene to a bare minimum by performing a fast simple segmentation into primary areas using spatial and/or quantization compression. (2) REFINE: resegment the scene with careful attention directed to the textural complexities of each region. The primitive transformations which are used include histogram clustering, region growing, data reduction by narrowing the quantization range, and/or data reduction by spatially collapsing the data while extracting features. These algorithms have been implemented using a parallel, hierarchical computational structure. Comparisons of performance on several images are given. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- May 01, 1977
- Accession Number
- ADA043676
Entities
People
- Allen R. Hanson
- Edward M. Riseman
- Paul A. Nagin
Organizations
- University of Massachusetts Amherst