Hierarchical Segmentation of Foveal Images
Abstract
This paper describes a technique for the segmentation of foveal images and other foveal retinotopic feature maps. The technique is implemented hierarchically in the small subset of an image pyramid, called the foveal polygon, supported by foveal imaging. Unlike binary or single spot detectors, this technique segments into multiple classes; the number of classes is determined by the image itself, and a maximum within-segment feature variance threshold. The technique represents segments, which need not be retinotopically contiguous, as subtrees in the polygon. These subtrees are used for efficient segment analysis and manipulation; the subtree nodes representing a single compact region are quickly identified and labeled, and statistics are efficiently computed at all levels of the hierarchy, including on segments and compact regions. Region segmentation is a fundamental component of almost every machine vision system. Hierarchical segmentation on the pyramid has been demonstrated to yield better results than conventional 2-D segmentation techniques (Hird89), and has been shown to quickly converge to a stable solution (Cibulskis84). This paper examines the hierarchical segmentation techniques developed for the foveal polygon. The basic segmentation technique was adapted from the multiclass pyramidal techniques developed by Burt and Rosenfeld (Burt8l) (Hong82). Several extensions have improved segmentation results, yielding multiple homogeneous, compact regions within segments. It has also been extended to work with foveal data.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 01, 1992
- Accession Number
- ADA282273
Entities
People
- Andrew Izatt
- Cesar Bandera