Mechanization of Context Information.
Abstract
This is a theoretical and experimental study of ways to use context for automatic interpretation of aerial photography. One type of photo interpretation system interprets a frame of aerial photography by dividing the frame into several square cells and classifying each cell. A parameter extraction device scans each cell, makes measurements, and classifies each cell according to statistical decision theory. System improvements normally require that new measurements be made and that the old measurements be refined. This study, however, addresses the use of contextual information from measurements made on neighboring cells. Since real measurement data was unavailable, measurements were simulated. The results of this study, however, have been shown to be independent of the particular simulation. A rule, for drawing contextual information about a cell from each of its four adjacent neighbors independently, has been derived from fairly general assumptions. Tests indicated that context can cut the error rate in half. Region theory has been developed to prevent the fixed cell size, dictated by hardware considerations, from interfering with the context mechanization. Although it has no experimental verification, calling a connected set of similar cells a region, inferring that these cells constitute one target, a recognizing frames of regions rather than cells, improves performance by using region geometry to characterize the target, and by drawing context only from the region boundary cells. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 01, 1969
- Accession Number
- AD0850026
Entities
People
- John R. Welch
- Kenneth G. Salter