Pattern recognition constraint network

• US 6,122,399 A
• Filed: 09/04/1997
• Issued: 09/19/2000
• Est. Priority Date: 09/04/1997
• Status: Expired due to Term

• Current Assignee: NCR Voyix Corporation
• Original Assignee: NCR Corporation
• Inventors: Moed, Michael C.
• Primary Examiner(s): Mancuso, Joseph
• Assistant Examiner(s): Bali, Vikkram
• Application Number:
  US08/923,504
• Time in Patent Office: 1,111 Days
• Field of Search: 382/156, 382/157, 382/159, 382/158, 382/225, 382/239, 706/20
• US Class Current: 382/159
• CPC Class Codes: G06F 18/24765 Rule-based classification
• CPC Class Codes: G06V 30/10 Character recognition
• CPC Class Codes: G06V 30/1444 Selective acquisition, locating or processing of specific regions, e.g. highlighted text, fiducial marks or predetermined fields
• CPC Class Codes: G06V 30/416 Extracting the logical structure, e.g. chapters, sections or page numbers; Identifying elements of the document, e.g. authors

Abstract

The present invention provides a novel pattern recognition system that may be used to determine the likelihood that an unknown item belongs to a predefined class. The present invention provides improved functionality over prior art systems in that both feature parameters and confidence parameters are trained automatically from ground truth image data; large numbers of different attributes can be analyzed in the model, which yields more accurate modeling; relationships between different features can also be trained from ground truth data and represented in parametric form; there is insensitivity to feature noise; parameters are based exclusively on statistical information; the need for heuristic information is minimized; retraining is straightforward and fast; on-line updating of model parameters is possible; and the models created are accurate and reliable at predicting the match of an unknown item to the model.

Claims

1. A pattern recognition system comprising:

(a) input means for receiving feature values of an unknown object, wherein the object belongs to one of a plurality of models, and wherein each feature value has a known range; and

(b) constraint means for providing a network of constraints representing one of the plurality of models, the constraint means comprising a plurality of weights, a plurality of nodes and a plurality of layers wherein each layer of nodes corresponds to a particular feature and each node is associated with a sub-range of the typical values for that feature for a selected one of the plurality of models, and wherein each weight corresponds to a value representing a relationship between the nodes of at least one feature type and the selected model, the constraint means performing the steps of;

(c) receiving the features from the input means; and

(d) responsive to the relationship between those nodes that correspond to the feature values and the weights between those nodes, generating a value corresponding to the confidence that the object belongs to the selected model.

2. The pattern recognition system of claim 1, wherein each node corresponds to the deviation of the feature value of the unknown object from the mean value for that feature.

3. The pattern recognition system of claim 2, wherein the plurality of nodes are ordered by distance from the mean value for that feature.

4. The pattern recognition system of claim 1, wherein each node corresponds to an ordered sub-range of values.

5. A pattern recognition process for generating a confidence level that an unknown object which belongs to one of a plurality of models actually belongs to a selected model, the pattern recognition process for use with a constraint network, the constraint network providing a network of constraints representing one of the plurality of models, the constraint network comprising a plurality of weights, a plurality of nodes and a plurality of layers wherein each layer of nodes corresponds to a particular feature and each node is associated with a sub-range of the typical values for that feature for a selected one of the plurality of models, and wherein each weight corresponds to a value representing a relationship between the nodes of at least one feature type and the selected model, the pattern recognition process comprising the steps of:

(a) receiving feature values of an unknown object, wherein each feature value has a known range; and

(b) responsive to the relationship between those nodes that correspond to the feature values and the weights between those nodes, generating a value corresponding to the confidence that the object belongs to the selected model.

6. The pattern recognition process of claim 5, wherein each node corresponds to the deviation of the feature value of the unknown object from the mean value for that feature.

7. The pattern recognition process of claim 6, wherein the plurality of nodes are ordered by distance from the mean value for that feature.

8. The pattern recognition process of claim 5, wherein each node corresponds to an ordered sub-range of values.