US 2014/0006334 A1Published application

System and method for scaled multinomial-dirichlet bayesian evidence fusion

AssigneeRaytheon Company

InventorDavid M. Doria

Life of this patent

Estimating the term…

Loading patent…

Abstract

A fusion method, implemented by one or more processors, for classifying a target having class types. The method includes: obtaining evidence from one or more classifiers, the evidence represented by scores from the one or more classifiers; representing the obtained evidence in a Bayesian context, where the Bayesian beliefs are obtained from the scores; obtaining new evidence, the new evidence represented by new scores from the classifiers; representing the obtained new evidence in an enhanced Bayesian context, where the enhanced Bayesian beliefs are obtained from the new scores; combining the scores and the new scores over multiple times; combining the evidence and the new evidence over multiple times; and using the combined scores and the combined evidence for each of the plurality of class types to classify the target.

Claims12

Dependent claims

A fusion method, implemented by one or more processors, for classifying a target having a plurality of class types, the method comprising:

obtaining evidence from one or more classifiers;
obtaining prior distribution of Bayesian beliefs of the class types of the target;
obtaining information about recognition accuracy and capacity of each of the one or more classifiers;
determining multinomial distribution model, based on the obtained evidence and the information about recognition accuracy and capacity;
representing the obtained prior distribution of Bayesian beliefs by a Dirichlet distribution model;
applying the Dirichlet distribution model to the determined multinomial distribution model to obtain an updated Dirichlet distribution model;
obtaining new evidence;
determining a new multinomial distribution model, based on the obtained new evidence and the information about recognition accuracy and capacity;
applying the updated Dirichlet distribution model to the new multinomial distribution model to obtain a new updated Dirichlet distribution model;
using parameters of the new updated Dirichlet distribution model to estimate Bayesian confidence values for each of the plurality of class types;
and
using the estimated Bayesian confidence values for each of the plurality of class types to classify the target.

9 more elements

Dependencies are inferred from the claim text.