Reliable identification with preselection and rejection class

Details Reliable identification with preselection and rejection class Reliable identification with preselection and rejection class

2000-02-07

2003-02-11

Starks, Wilbert L. (Department: 2121)

Data processing: artificial intelligence

Knowledge processing system

C706S020000, C382S170000, C382S226000

Reexamination Certificate

active

06519579

ABSTRACT:

A classification of data sets (e. g. picture data, speech signals) is the basis of an “intelligent” computer performance. Numerous fields of use exist, e. g. industrial production, biometrical recognition of humans, medical picture processing, etc.
The state of the art comprises a great number of classificators, e. g.
statistical classificators (Gaussian distribution classificators)
neuronic networks
synergistic algorithms
next-neighbor classificator.
A standard literature in the field of pattern recognition is Nieman, “Klassifikation von Mustern”, Springer Verlag, 1983.
It is an object of the present invention to further improve classification quality by providing new classificators or new basic formulations of a classification.
This object is solved by one of claim 1 and 10.
Explanations are intended to contribute to a better comprehension of the technical terms used in the claims.
Identification (or Classification) of n Classes:
After providing n classes from a predetermined representative off-hand sample in a so-called learning process, an association/classification of a (still) unknown pattern into a certain class is called ‘identification’. By introducing a rejection threshold, a pattern may be rejected as unknown. If it approximates a rejection class more closely than one of learned and known target classes of said identification, it is classified into said rejection class. A rejection threshold and a rejection class may be provided alternatively and cumulatively. A pattern is regarded as a “rejected pattern” (object or person), if all “rejections” provided (at least one of a threshold and a class) have responded. A precondition for a successful “identification” is that the test pattern provides sufficient information for a clear association to one of said n classes of the learning process.
Verification:
An identification with n=1 is made by an a priori (previous) knowledge concerning the target class, i. e. like a binary decision, (only) an acceptance or a rejection of the test pattern (patterns used for the test, shortly: “test pattern”) may result.
FAR, FRR, Quality Function:
FAR (false acceptance rate) designates the rate of patterns identified false; FRR (false rejection rate) designates the rate of patterns rejected false. A quality function G=G (FAR,FRR) indicates the quality of a classification process, e. g. G=1−FAR−FRR. The more precise a classification, the closer G approaches “one”. A weighting of FAR and FRR may have an influence if one or the other parameter FAR, FRR shall be accentuated, e. g. by indicating an average value with weighting factors g
1
,g
2
, e. g. (g
1
·FAR+g
2
·FRR)/(g
1
+g
2
). In practical applications, FRR may be of more importance, so that e. g. g
2
=2 and g
1
=1 may be selected to make said quality G “measurable” and to be able to compare identifications.
The method according to the invention serves to improve classification quality.
(a) In a first step, an ‘identification’ of n classes is provided, said identification being improved by a double use of a provided information. For this purpose, an information content of a test pattern is split into a necessary and a sufficient portion to be associated to a class. With said necessary portion, a preselection (pre-classification) of the classes to be considered may be effected. With said method, no precise (but rather an imprecise) classification is obtained, but the number of classes to be really considered for the pattern is substantially limited or reduced. Said step provides a “better” identification (in the meaning of the above quality function G).
(b) In a second step, classification quality is improved by an (additional) rejection class. Said class serves to support a rejection, i. e. in addition to rejections obtained for instance by threshold decisions, a particular rejection class is equally entitled with respect to the identification classes (the concrete target classes), into which a classification may be effected. With said rejection class, an a priori (previous) knowledge concerning the objects/persons (general term: patterns) to be rejected is taken into consideration insofar as e. g. a representative profile of the “patterns” to be rejected is learned into said rejection class and is therefore known to the classificator.
The composition of the rejection class is “subject to success”, i. e. (in the meaning of the quality function) a better solution of the classification problem has to be provided by using said rejection class. When patterns to be rejected are e. g. known, they may all be learned into said rejection class. But usually, only a certain portion thereof is necessary for an improved rejection class, another portion may for example originate from a data base not relating to this problem at all.
For the selection of patterns to be rejected, selection methods may be used, e. g. testing all possibilities and using those data sets (of patterns) which yield the best result.


REFERENCES:
patent: 4166540 (1979-09-01), Marshall
patent: 5329596 (1994-07-01), Sakou et al.
patent: 5359699 (1994-10-01), Tong et al.
patent: 5987170 (1999-11-01), Yamamoto et al.
patent: 38 34 869 (1996-08-01), None
Neimann, Heinrich: “Information Processing in Technical, Biological and Economic Systems: Methods of Pattern Recognition”Akademische Verlagsgesellschaft, Frankfurt am Main 1974, pp. 410-414, portion 4.3.2.3.
Schurmann J: “Zur Zuruckweisung zweifelhafter Zeichen”Nachrichtentechnische Zeitschrift, Mar. 1973, West Germany, vol. 26, No. 3, pp. 137-144, XP002081339 ISSN 0027-707X.
Schwerdtmann W: “Reduktion Des Klassifikationsfehlers Durch Angepasste Musterentwicklung”NTZ Nachrichtentechnische Zeitschrift, vol. 27, No. 6, 1974, pp. 233-238, XP002067426.
Fujimoto Y et al: “Recognition of Handprinted Characters by Nonlinear Elastic Matching”3rd International Joint conference on Pattern Recognition, Coronada, Ca, USA, Nov. 8-11, 1976, pp. 113-118, XP002081340 1976, New York, NY, USA, IEEE.
Fleming M.K. et al: “Categorization of Faces Using Unsupervised Feature Extraction”International Joint Conference on Neural Networks(IJCNN), San Diego, Jun. 17-21, 1990, vol. 2, Jun. 17, 1990, pp. 65-70, XP000144325 Institute of Electrical and Electronics Engineers.
Wagner T. et al: “Sensor-fusion for Robust Identification of Persons: A field Test”Proceedings of the International Conference on Image Processing(ICIP), Washington, Oct. 23-36, 1995, vol. 3, Oct. 23, 1995, pp. 516-519, XP000623197 Institute of Electrical and Electronics Engineers, para. 2, Synergetic Computer as a Classifier.
Dieckmann U. et al: “SESAM: a biometric person identification system using sensor fusion”Audio- and Video-Based Biometric Person Authentication. First International Conference, AVBPA'97, Proceedings, Proceedings of First International Conference on Audio and Video Based Biometric Person Authentication (AVBPA), 301-310, XP002081341, ISBN 3-540-62660-3, 1997, Berlin, Germany, published by Verlag, Germany.

Affiliated with

Also associated with

Rating

Reliable identification with preselection and rejection class does not yet have a rating. At this time, there are no reviews or comments for this patent.

If you have personal experience with Reliable identification with preselection and rejection class, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Reliable identification with preselection and rejection class will most certainly appreciate the feedback.

Rate now

Comments { 0 }

Profile ID: LFUS-PAI-O-3137981