Image analysis – Learning systems – Trainable classifiers or pattern recognizers
Patent
1997-09-08
1999-08-24
Au, Amelia
Image analysis
Learning systems
Trainable classifiers or pattern recognizers
G06K 962
Patent
active
059434384
DESCRIPTION:
BRIEF SUMMARY
BACKGROUND OF THE INVENTION
1. Field of Invention
The present invention relates to a method for recognizing at least one defined pattern modeled by hidden Markov models in a time-variant measurement signal on which at least one disturbing signal is superposed.
2. Description of the Related Art
Pattern recognition is gaining in importance in technical systems, due to increased use of automatization. The patterns are often contained in measurement signals obtained in one way or another from processes that are to be examined. Examples include the analysis of natural language or the examination of executed signatures. Hidden Markov models have proven particularly suitable for the automatic analysis of measurement signals. For example, see the publication by Rabiner, et al, "An Introduction to Hidden Markov Models".
However, a problem in the automatic analysis of measurement signals is that the measurement signals to be examined are often overlaid with variable or occasional disturbing signals or with disturbing signals that are constant in quantity. In the acquisition and analysis of naturally uttered speech, these can be e.g. background noises, breathing noises, machine noises, or also disturbing noises that arise due to the recording medium and the transmission path. For other measurement signals, analogous error sources are possible. In order to be able to find a known pattern within a larger pattern despite the existing difficulties, an approach with special hidden Markov models has been proposed in the publication by Rose et al., "A Hidden Markov Model Based Keyboard Recognition System". A specific model is hereby introduced (termed a garbage or background model) that models background noise, including other speech. This specific model must in all cases be trained with corresponding noises or, respectively, speech. This means that only those noises or, respectively, disturbances that are present in the training material can be taken into account when recognizing a pattern. In each case, the modeling of this model has a large influence on the overall probability of a key word, or, of any pattern to be recognized. A further disadvantage is that this background model also recognizes speech sounds, or, respectively, pattern portions in other types of measurement signal, which actually belong to a pattern or, respectively, a keyword to be recognized. The additional problem of a suitable weighting of the model in order to avoid an excessively low detection rate thereby results, as discussed in the Rose publication. Further possibilities for taking into account disturbing portions in the analysis of measurement signals, using hidden Markov models, are not known from the prior art.
SUMMARY OF THE INVENTION
The underlying aim of the present invention is thus to provide a method with which patterns that are present in a measurement signal, and on which a disturbing signal is superposed, can better be recognized.
This aim is solved by the method for recognizing at least one defined pattern modeled by hidden Markov models in a time-variant measurement signal on which at least one disturbing signal is superposed, beginning and at its end by respectively one single identical state that serves to represent the disturbing signal and has least the following characteristics: pattern, it comprises no transition probabilities, environment of the defined pattern, it comprises no emission probabilities, comparison method, using the expand hidden Markov models in the time-variant measurement signal.
Developments of the invention are provided by the method in which the Viterbi algorithm is used as a comparison method. In the alternative, the Baum-Welch method is used as a comparison method. In either case, in the comparison method in each recombination of paths, the overall emission probabilities accumulated per path are normed to the number of the emission probabilities taken into account up to this time, in order to avoid a preferring of shorter paths.
The norming ensues in the following way: ##EQU1## Segm(l)=Segm(N)=S.sub
REFERENCES:
patent: 5502791 (1996-03-01), Nishimura et al.
patent: 5812974 (1998-09-01), Hemphill et al.
Y. Ephraim et al., "On The Application Of Hidden Markov Models For Enhancing Noisy Speech", IEEE Transactions on Acoustics, Speech, and Signal Processing, vol. 37, No. 12, Dec. 1989, pp. 1846-1856.
J. Song, "A Robust Speaker-Independent Isolated Word HMM Recognizer For Operation Over The Telephone Network", 8308 Speech Communication, Dec. 1993, Nos. 3/4, pp. 287-295.
R. Cardin et al., "Inter-Word Coarticulation Modeling and MMIE Training for Improved Connected Digit Recognition", Speech Processing, Apr. 1993, pp. II-243-II-246.
L.R. Rabiner et al., "An Introduction to Hidden Markov Models", IEEE Transactions of Acoustics, Speech and Signal Processing, 1986, pp. 4-16.
G. David Forney, Jr., "The Viterbi Algorithm", Proceedings of the IEEE, vol. 61, No. 3, Mar. 1973, pp. 268-278.
R. Rose et al., A Hidden Markov Model Based Keyword Recognition System.sup.1, IEEE International Conference on Acoustics, Speech and Signal Processing, 1990, pp. 129-132.
Ahmed Samir
Au Amelia
Siemens Aktiengesellschaft
LandOfFree
Method for pattern recognition does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Method for pattern recognition, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method for pattern recognition will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-474472