KBVision Constraint System Module Page 1
KBVision Constraint Module
Overview
The Constraint System is a general purpose interactive tool that is used to create discriminative functions whose inputs are values of extracted Features, and whose outputs are levels of confidence in the existence of some object. Confidence levels can be used for:
- Incorporation of domain or user knowledge to guide further processing or to aid in object classification
- Grouping, feature correlations, and statistical information for Token-based data, such as lines, regions, surfaces, and other Image or data related events
- Determination of areas of interest for intelligent pre-processing of associated low-level (pixel-based) Image representations
What Is A Constraint?
A Constraint is a mathematical function that defines the mapping from one or more Token feature values (attributes) to a Constraint score. A Constraint Set is simply a set of Constraints and a Property List.
Constraint scores are generally used to numerically symbolize the presence and magnitude of characteristics of domain objects. Often, this becomes a measure of evidence or confidence that a particular object is present at the Token(s) location, and can be used to classify Tokens, or act as initial hypothesis values for later processing.
Once a Constraint has been displayed for a Tokenset, a Constraint score for every Token is temporarily included as part of the Tokenset as a Token feature. The Token feature name is identical to the Constraint name, and the Token feature value is the constraint score.
The option exists in the Constraint System to save the Tokenset with the newly-defined Constraint-based features. In addition, once the Constraint results become part of the Tokenset, they can be used as features to form other Constraints. This Constraint hierarchy is possible whether the Tokenset is saved with the Constraint scores or not.
Types of Constraints
- Expression - Expression Constraints specify functions that mathematically combine Token feature values.
- Primitive Constraints - Primitive Constraints are Piece-wise linear functions that mapping Token feature values onto Constraint scores. Histograms can be used to visually guide the selection feature of value-score pairs.
- Lookup - Lookup Constraints are identical to Primitive Constraints, but take on a different internal representation. They can be easily downloaded to lookup tables to seed up processing.
- Compound - Compound constraints are functions that combine the resulting scores from other Constraint functions.
- Relational - Relational Constraints combine Constraints that require Token information from two Tokensets.
As an aid in the construction of Constraints that directly map feature values to scores, (Primitive and Lookup), it is possible to manually create Tokensets that can be included in the histogram display. This is accomplished by explicitly selecting Tokens from a display of the outlines of all Tokens in the tokenset of the Image representation. Each of these Tokensets has an associated display color. When a histogram is displayed, bins representing Tokens from a Tokensubset will be displayed in the Tokensubset color. This display is very useful in extracting initial feature correlations for objects of interest.
It is also possible to display a combined histogram of more than one Tokenset. This is especially valuable to ensure that Constraints do not become too specialized, such that they describe only the specific appearance of an object in a single Image.
CONSTRAINT SYSTEM Page 1 of 4
Go to next Page
KBVision Modules
webmaster@aai.com (Last Update 10/17/95)
KBVision, KBView are Trademarks of Amerinex Artificial Intelligence, Inc.