The Design of a Semantically Directed Vision Processor (Revised and Updated).
Abstract
This paper updates the design of a semantically directed vision processor. The system will carry out model-directed analysis of outdoor scenes by applying semantic knowledge at an early stage of processing. The goal is to quickly and flexibly interface low-level visual features (e.g., edge detectors, texture and color analyses) and high-level conceptual knowledge (e.g., trees stem from the ground, general knowledge associated with road scenes, and the like) in the perception of complex images. The computational structure for rapidly extracting visual features is called a 'processing cone.' The cone consists of parallel spatial arrays of micro-computing elements, each of which operate on a local window to reduce the data layer by layer. Information flows up, down, and laterally in the cone via a sequence of local parallel operations. Routines for detecting objects will examine the data at the top of the cone and will selectively analyze the lower level mass of data. Rough confidences of the presence of objects in various regions will be passed to the model builder.
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 01, 1975
- Accession Number
- ADA010150
Entities
People
- Allen R. Hanson
- Edward M. Riseman
Organizations
- University of Massachusetts Amherst