Automatic classifier for real time inspection and classification

• US 6,092,059 A
• Filed: 12/27/1996
• Issued: 07/18/2000
• Est. Priority Date: 12/27/1996
• Status: Expired due to Term

• Current Assignee: Ametek Incorporated
• Original Assignee: Cognex Corporation
• Inventors: Lavagnino, Sherrill E., Straforini, Marco L., Tilson, Bret R., Wolinsky, Jeffrey M., Badger, John C.
• Primary Examiner(s): Downs, Robert W.
• Assistant Examiner(s): Shah, Sanjiv
• Application Number:
  US08/777,367
• Time in Patent Office: 1,299 Days
• Field of Search: 706/14, 706/20, 706/10, 706/47, 706/12, 382/157, 382/158, 382/159, 328/128, 378/37
• US Class Current: 706/14
• CPC Class Codes: G06F 18/254 of classification results, e.g. of results related to same input data

Abstract

An improved classification system for classifying characteristics of an article into classes of characteristics based on features associated with the characteristics. The system includes a first rule-based classifier that provides rules that embody requirements of selected features so as to together define selected classes. These rules are applied to measured features of an article characteristic to assign one of the selected classes to that characteristic, should the article characteristic meet the feature requirements of one of the selected classes. The system also includes a trainable classifier configured based on a training set of preclassified characteristics, with the trainable classifier connected so as to accept from the first rule-based classifier an article characteristic not classified by the rule based classifier. The trainable classifier assigns one of the training set classes to the unclassified characteristic, should the trainable classifier configuration associate that characteristic with one of the training set classes. A second rule-based classifier can be connected so as to accept classified characteristics from the trainable classifier for making refinement classifications, and a third rule-based classifier can be connected so as to accept unclassified characteristics from the trainable classifier for making classification based on selected classes.

Claims

1. A classification system for classifying the characteristics of an article into classes of characteristics based on features associaated with the characteristics, the system comprising:

a first rule-based classifier providing rules that embody requirements of selected features so as to together define selected classes, the rules to be applied to measured features of an article characteristic to assign one of the selected classes to that characteristic, should the article characteristic meet the feature requirements of one of the selected classes; and

a trainable classifier that is configured based on a training set of preclassified characteristics and a corresponding training class set which can include characteristic classes that are not the selected classses of the first-rule based classifier, the trainable classifier connected so as to accept from the first rule-based classifier an article characteristic not classified by the first ruled-based classifier, to assign one of the training set classes to thea characteristic, should the trainable classifier configuration associate that characteristic with one of the training set classes.

6. The classification system of claim 1 wherein the selected classes defined by rules of the first rule-based classifier comprise classes expected to be frequently assigned to article characteristics of interest and classes for which assignment to characteristics of interest, based on rigid feature requirements, is expected to be accomplished with at least a minimum prespecified level of confidence.

7. The classification system of any of claims 1, 2, or 3, further comprising means for choosing the features of the preclassified characteristic training set based on a comparison of percentage of training set characteristics that are associated by the trainable classifier with the class known to be correct for each respective training set characteristic, the comparison carried out as each possible feature is added to a set of possible features, each feature being added in an order beginning with a feature known to provide a highest degree of correctness in association by the trainable classifier with the characterisitic training set, based on that feature, any possible feature for which comparison of the association percentage of the training set using a current feature accumulation in the set of possible features is less than an association percentage for a just preceeding feature accumulation in the set being indicated as not to be chosen as a feature of the preclassified characteristic training set.

8. The classification system of any of claims 1, 2, or 3, further comprising means for generating the measured features of characteristics of the article.

9. The classification system of claim 8 further comprising means for acquiring characteristic information from the article to be examined for generating the measured features.

10. The classification system of any of claims 1, 2, or 3, wherein the rules provided by the first rule-based classifier are configured as a classification table, with each classification table row corresponding to a selected class and each column of a given class row corresponding to a feature requirement for that class.

13. The classification system of any of claims 1, 2, or 3, wherein the trainable classifier is configured based on probability functions that are constructed based on probability of association by the trainable classifier of each characteristic of the preclassified characteristic training set with a corresponding known respective class for that characteristic.

14. The classification system of claim 13 wherein at least one feature of characteristics in the training set is weighted for each class in the training set as a function of degree of criticality of that feature in correct association of a characteristic with that class.

2. A classification system for classifying charateristics of an article into classes of characteristics based on features associated with the characteristics, the system comprising:

a trainable classifier configured based on a training set of preclassified characteristics for assigning one of the training set classes to an article characteristic, should the trainable classifier configuration associate that characteristic with one of the training set classes; and

a first rule-based classifier connected so as to accept from the trainable classier a classified article characteristic for refining classification of that characteristic, the first rule-based classifier providing refinement rules embodying requirements of selected refinement features so as to together define refinement classes, the refinement rules to be applied to measured features of the classified article characteristic for assignment one of the refinement classes to the classified characteristic, should the classified characteristic meet the refinement feature requirements of one of the refinement classes.

15. The classification system of either of claims 2 or 5, wherein the refinement characteristic classes comprise characteristic subclasses of at least one training set class of the trainable classifier.

16. The classification system of either of claims 2 or 5, wherein the refinement characteristic classes comprise classes defined by confirmational feature requirements for confirming validity of trainable classifier class associations.

17. The classification system of either of claims 2 or 5, wherein the refinement characteristic classes comprise classes defined by clarification feature requirements for clarifying a trainable classifier class association between two classes for which confusion between the two classes is possible.

19. The classification system of claim 2 further comprising:

a second rule-based classifier connected so as to accept from the trainable classifier an article characteristic not classified by the trainable classifier, the second rule-based classifier providing rules that embody requirements of selected features to as to together define selected classes, the rules to be applied to measured features of the unclassified article characteristics for assigning one of the selected classes to the unclassified characteristic, should the unclassified article characteristic meet the feature requirements of one of the selected classes.

3. A classification system for classifying characteristics of an article into classes of characteristics based on features associated with the characteristics, the system comprising:

a trainable classifier configured based on a training set of preclassified characteristics for assigning one of the training set classes to an article characteristic, should the trainable classifier configuration associate that characteristic with one of the training set classes; and

a first rule-based classifier connected so as to accept from the trainable classifier an article characteristic not classified by the trainale classifier, the first rule-based classifier providing rules that embody requirements of selected features to as to together define selected classes, the rules to be applied to measured features of the unclassified article characteristic for assigning one of the selected classes to the unclassified characteristic, should the unclassified article characteristic meet the feature requirements of one of the selected classes.

4. A classification system for classifying characteristics of an article into classes of characteristics based on features associated with the characteristics, the system comprising:

a first rule-based classifier providing rules that embody requirements of selected features so as to together define selected classes, the rules to be applied to measured features of an article characteristic to assign one of the selected classes to that characteristic, should the article characteristic meet the feature requirements of one of the selected classes;

a trainable classifier configured based on a training set of preclassified characteristics, the trainable classifier connected so as to accept from the first rule-based classifier an article characteristic not classified by the rule based classifier, to assign one of the training set classes to that characteristic, should the trainable classifier configuration associate that characteristic with one of the training set classes; and

a second rule-based classifier connected so as to accept from the trainable classifier an article characteristic not classified by the trainable classifier and the first rule-based classifier, the second rule-based classifier providing rules that embody requirements of selected features so as to together define selected classes, to be applied to measured features of the unclassified article characteristic for assigning to the unclassified characteristic one of the selected classes, should the unclassified characteristic meet the feature requirements of one of the selected classes.

5. A classification system for classifying characteristics of an article into classes of characteristics based on features associated with the characteristics, the system comprising:

a first-rule based classifier providing rules that embody requirements of selected features so as to together define selected classes, the rules to be applied to measured features of an article characteristic to assign one of the selected classes to that characteristic, should the article characteristic meet the feature requirements of one of the selected classes;

a trainable classifier configured based on a training set of preclassified characteristics, the trainable classifier connected so as to accept from the first rule-based classifier an article characteristic not classified by the rule based classifier, to assign one of the training set classes to that characteristic, should the trainable classifier configuration associate that characteristic with one of the training set classes; and

a second rule-based classifier connected so as to accept from the trainable classifier a classified article characteristic for refining classification of that characteristic, the second rule-based classifier providing refinement rules that embody requirements of selected refinement features so as to define refinement classes, the refinement rules to be applied to measured features of the classified article characteristic to assign one of the refinement classes to the classified characteristic, should the classified characteristic meet the refinement feature requirements of one of the refinement classes.

11. The classification system of claim 5 further comprising a third rule-based classifier connected so as to accept from the trainable classifier an article characteristic not classified by the trainable classifier and the first rule-based classifier, the third rule-based classifier providing rules that embody requirements of selected features, so as to together define selected classes, the rules to be applied to measured features of the unclassified article characteristic for assigning one of the selected classes to the unclassified characteristic, should the unclassified characteristic meet the feature requirements of one of the selected classes.

12. The classification system of any of claims 4, 11, or 3, wherein a plurality of rules provided by the rule-based classifier that is connected so as to accept from the trainable classifier an unclassified article characteristic relate to a prespecified minimum degree of confidence of classificaiton assignment by the trainable classifier, one degree of confidence of each of a corresponding plurality of selected classes, a corresponding minimum degree of confidence required for assignment of a given one of the plurality of selected classes to the unclassified article characteristic.

18. The classification system of any of claims 4, 11 or 3, wherein the rules provided by the rule-based classifier that is connected so as to accept an unclassified article from the trainable classifier are configured as a classification table, with each classification table row corresponding to a selected class and each column of a given class row corresponding to a feature requirement for that class, the classes selected for enabling classification of article characteristics undefined by another classifier in the classification system.

20. An image classification system comprising:

feature extraction means responsive to an image of an article to produce feature information about characteristics of the article for classification of the characteristics into classes;

a first rule-based classifier providing rules that embody requirements of selected features so as to together define selected classes, the rules to be applied to the extracted feature information of a given characteristic for assigning one of the selected classes to that characteristic, should the article characteristic meet the feature requirements of one of the selected classes;

a trainable classifier configured based on a training set of preclassified characteristics, the trainable classifier connected so as to accept from the first rule-based classifier an article characteristic not classified by the rule based classifier, to assign one of the training set classes to that characteristic, should the trainable classifier configuration associate that characteristic with one of the training set classes;

a second rule-based classifier connected so as to accept from the trainable classifier a classified article characteristic for refining classification of that characteristic, the second rule-based classifier providing refinement rules that embody requirements of selected refinement features so as to together define refinement classes, the refinement rules to be applied to extracted features of the classified article characteristic to assign one of the refinement classes to the classified characteristic, should the classified characteristic meet the refinement feature requirements of one of the refinement classes; and

a third rule-based classifer connected so as to accept form the trainable classifier an article characteristic not classified by the trainable classifier and the first rule-based classifier, the third rule-based classifier providing rules that embody requirements of selected features so as to together define selected classes, the rules to be applied to extracted features of the unclassified article characteristic to assign one of the selected classes to that characteristic, should the article characteristic meet the feature requirements of one of the selected classes.

21. A method for classifying characteristics of an article into classes of characteristics based on features associated with the characteristics, the method comprising:

applying rules, that embody requirements of selected features so as to define selected classes, to measured features of an article characteristic to be classified;

assigning one of the selected classes to the article characteristic, should the characteristic meet the feature requirements of the rules for one of the selected classes; and

providing a characteristic that is not assigned one of the selected classes to the trainable classifier that is configured based on a training set of preclassified characteristics and a corresponding training class set which can include characteristic classes that are not the selected classes, the trainable classifier assigning one of the training set classes to the unclassified characteristic, should the trainable classifier configuration associate that characteristic with one of the training set classes.

22. A method for classifying characteristics of an article into classes of characteristics based on features associated with the characteristics, the method comprising:

applying rules, that embody requirements of selected features so as to define selected classes, to measured features of an article characteristic to be classified;

assigning one of the selected classes to the article characteristic, should the characteristic meet the feature requirements of the rules for one of the selected classes;

providing a characteristic that is not assigned one of the selected classes to a trainable classifier that is configured based on a training set of preclassified characteristics, the trainable classifier assigning one of the training set classes to the unclassified characteristic, should the trainable classifier configuration associate that characteristic with one of the training set classes;

accepting an unclassified article characteristic from the trainable classifier;

applying to measured features of the unclassified characteristic rules that embody requirements of selected features so as to together define selected classes; and

assigning one of the selected classes to the unclassified characteristic, should the unclassified characteristic meet feature requirements of the rules for one of the selected classes.

23. A method for classifying characteristics of an article into classes of characteristics based on features associated with the characteristics, the method comprising:

applying rules, that embody requirements of selected features so as to define selected classes, to measured features of an article characteristic to be classified;

assigning one of the selected classes to the article characteristic, should the characteristic meet the feature requirements of the rules for one of the selected classes;

providing a characteristic that is not assigned one of the selected classes to a trainable classifier that is configured based on a training set of preclassified characteristics, the trainable classifier assigning one of the training set classes to the unclassifed characteristic, should the trainable classifier configuration associate that characteristic with one of the training set classes;

accepting a classifed article characteristic from the trainable classifier;

applying to measured features of the classified characteristic refinement rules that embody requirements of refinement features so as to together define refinement classes; and

assigning one of the refinement classes to the classified characteristic, should the classified characteristic meet the refinement feature requirements of the refinement rules for one of the refinement classes.