Multiresolutional classifier with training system and method
AssigneeMotorola, Inc.
InventorsKhaled Assaleh, William Michael Campbell, John Eric Kleider
Life of this patent
Estimating the term…
Abstract
An improved method and system for training and classifying using a low complexity and high accuracy multiresolutional polynomial classifier (412) is presented. A method of training an multiresolutional polynomial classifier which reduces the complexity of existing classifiers allows models representing subgroups of classes to easily be created. The models which represent subgroups of classes are applied to an unidentified input to produce a coarse classification of the unidentified input using a low order classifier. Once a coarse classification of the unidentified input is performed, a more detailed classification is performed using another low complexity classifier.
Claims14
A method of training a multiresolutional polynomial classifier (MPC) in a multiresolutional polynomial classifier and training system (MPCTS), wherein said MPCTS comprises said MPC, a superclass model database, a training processor, and a feature vector database, said MPC comprising a hierarchy of superclass classifiers, a superclass classifier comprising a model multiplier for each class in the subgroup, the method performed by said training processor operating on a group of superclass models for a plurality of classes, each of said plurality of classes being represented by a set of feature vectors stored in said feature vector database, each of said group of superclass models representing a superclass from a set of superclasses, each of said set of superclasses representing at least one class of said plurality of classes, said method comprising the steps of:
- a) determining a set of feature vectors representing a class of said plurality of classes;
- b) performing a linear predictive (LP) analysis for each feature vector of said set, said LP analysis generating a predetermined number of LP coefficients based on LP order, a feature vector being represented by a set of said LP coefficients;
- c) creating an individual class structure for each of said plurality of classes, said individual class structure being determined by summing feature vectors;
- d) determining if additional classes are part of a superclass, and when additional classes are part of said superclass performing steps a)-d);
- e) when additional classes are not part of said superclass, determining a superclass structure based on individual class structures, wherein individual class structures are combined to create a plurality of superclass structures;
- f) determining if additional superclasses require processing, and when additional superclasses require processing, performing steps a)-f);
- g) when additional superclasses do not require processing, determining a total class structure representing the set of classes, wherein said plurality of superclass structures are combined to create said total class structure;
- h) adding said total class structure to a scaled version of each of said plurality of superclass structures to create a combined class structure, ri,combined, for each of said set of superclasses, said combined class structure being mapped to a matrix;
- i) determining a superclass model for each of said set of superclasses using said combined class structure for each of said set of superclasses;
- and
- j) storing the superclass models.
A method of classifying an unidentified class as a member of a superclass, the method performed by a multiresolutional polynomial classifier and training system (MPCTS), wherein said MPCTS comprises a multiresolutional polynomial classifier (MPC), a superclass model database, a training processor, and a feature vector database, said superclass being represented by a superclass model, said superclass model being one of a group of superclass models in said superclass model database, said method comprising the steps of:
- a) determining a plurality of unidentified feature vectors representing said unidentified class, wherein said feature vector database comprises feature vectors;
- b) determining an average of the polynomial basis terms representing said unidentified class, said plurality of unidentified feature vectors being used to determine the average of the polynomial basis terms, wherein a summation is performed using the polynomial basis terms for each feature vector representing the unidentified class to determine a polynomial basis vector and the average of the polynomial basis vector is determined by dividing each basis term of the polynomial basis vector by the total number of feature vectors representing said unidentified class;
- c) identifying a node in a hierarchy of nodes, wherein at least one superclass model is identified for said node, wherein the superclasses which represent a node are preferably predetermined during a training procedure;
- d) classifying said unidentified class as a member of a superclass by multiplying a superclass model for each superclass represented by said node with the average polynomial basis vector, wherein a set of scalar values is established, said set of scalar values including one scalar value per superclass;
- e) selecting the largest scalar value from said set of scalar values, the largest score value representing a superclass to which said unidentified class is most likely a member;
- f) determining if another level of classification is needed, and when another level of classification is needed performing steps c)-f), wherein additional levels of classifications are performed at successive nodes in said hierarchy of nodes until said unidentified class is determined to be a member of at least one superclass and wherein a thread of nodes within said hierarchy of nodes is used to attain a desired level of classification;
- and
- g) when another level of classification is not needed, identifying said unidentified class as a member of said superclass, said unidentified class being classified as a member of the superclass associated with the largest scalar value.
A multiresolutional polynomial classifier and training system (MPCTS) for classifying an unidentified class as a member of a superclass from a set of superclasses, said unidentified class being represented by a plurality of unidentified feature vectors, said superclass being represented by a superclass model, said superclass model being one of a group of superclass models, said superclass models representing a plurality of classes, said system comprising:
- a multiresolutional polynomial classifier (MPC), said MPC comprising a plurality of superclass classifiers configured as a hierarchy of nodes, wherein said hierarchy of nodes are coupled to each other by at least one bus, thereby allowing said MPC to perform multi-level classifications of unidentified classes;
- a superclass model database coupled to said MPC via a bus, said superclass model database supplying feature vectors to model multipliers within each superclass classifier;
- a training processor coupled to said superclass model database, said training processor determining superclass models and providing said superclass models to said superclass model database;
- a feature vector database coupled to said training processor, said feature vector database for storing said plurality of unidentified feature vectors, wherein said training processor processes feature vectors obtained from said feature vector database to produce said superclass models;
- a feature input coupled to said MPC, said training processor, and said feature vector database, said feature input supplying feature vectors to said MPC;
- and
- a selector output coupled to said MPC, said MPCTS comprising means for training said MPC, and means for classifying an unidentified class as a member of a superclass based on superclass models created during said training, wherein the superclass which represents the unidentified class is output via said selector output, when the level of classification for the unidentified class is determined, said MPC further comprising means for averaging said plurality of unidentified feature vectors to produce an average command structure for said unidentified class, wherein said MPC performs a dot product with said group of superclass models and said average command structure to create a set of score values, each score value being associated with at least one superclass, and wherein said MPC identifies said unidentified class as said member of said superclass based on said set of score values.
Dependencies are inferred from the claim text.