EP2174267A2 - Procédé et dispositif de reconnaissance automatique de motifs - Google Patents
Procédé et dispositif de reconnaissance automatique de motifsInfo
- Publication number
- EP2174267A2 EP2174267A2 EP08801094A EP08801094A EP2174267A2 EP 2174267 A2 EP2174267 A2 EP 2174267A2 EP 08801094 A EP08801094 A EP 08801094A EP 08801094 A EP08801094 A EP 08801094A EP 2174267 A2 EP2174267 A2 EP 2174267A2
- Authority
- EP
- European Patent Office
- Prior art keywords
- data
- sequence
- model data
- electronic data
- sequences
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Withdrawn
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
- G06V10/761—Proximity, similarity or dissimilarity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/29—Graphical models, e.g. Bayesian networks
- G06F18/295—Markov models or related models, e.g. semi-Markov models; Markov random fields; Networks embedding Markov models
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/762—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using clustering, e.g. of similar faces in social networks
- G06V10/763—Non-hierarchical techniques, e.g. based on statistics of modelling distributions
Definitions
- the invention relates to a method and a device for automatic pattern recognition in a sequence of electronic data by means of electronic data processing in a data processing system.
- the goal of such pattern recognition is to find out the occurrence of sequences or sequences of properties in sequentially formed electronic data.
- the patterns to be found are not exactly definable because they may vary in shape and extent.
- An example is the problem of machine-language recognition, since essential standard methods of the prior art have been developed in the context of this task.
- Another application is finding fault patterns in machine signals. For example, the detection of knocking burns in gasoline engines by means of structure-borne noise signals, in which a very similar problem is involved (Lachmann et al .: Detecting knocking burns from disturbed knock sensor signals by means of signal separation, Sensory in the motor vehicle, Expert Verlag, 114 -123).
- HMM hidden Markov models
- a problem here is that the pattern sequences or sequences tend to vary in length, with two differently sized pattern sequences or sequences belonging to the same class. Sequences are thus not vectors, that is, there is no feature space and no probability distribution can be determined. The use of feature-vector-based classifiers is thus prevented.
- HMMs are parametric models, that is, they provide a limiting framework that is not always must fit. Parametric models are therefore often affected simultaneously by under- and over-adaptation to the example data.
- HMMs basically require the Markov property to be satisfied.
- Another example is the assumption of temporal invariance within a state. As a rule, both assumptions are never fully fulfilled, which results in a fundamental structural underfitting.
- a pattern recognition method which deals with the recognition of feature sequences - concretely with the recognition of speech - is described in DE 697 11 392 T2.
- Another field of application of pattern recognition of feature sequences or sequences relates to knock detection in connection with motors. This will be discussed in more detail below.
- Tapping burns are unwanted deviations from normal combustion. Normal burns are triggered by the spark plug spark and are associated with a moderate pressure increase in the cylinder. Knocking burns, however, generate high pressure peaks and can thus lead to damage to the engine. They often occur when the ignition occurs too early. A later ignition can remedy, but leads to a reduction in engine performance, and thus to an increase in fuel consumption. It therefore makes sense to choose the ignition timing so that just no knocking occurs. Since the Klopfhe Trent an engine depends on external influences, a knock-dependent control of the ignition timing is required. A reliable detection of knocking burns is indispensable.
- a knocking combustion can be determined on the basis of the pressure curve in the interior of the cylinder.
- sensors for detecting this measurement are expensive and wear out quickly, so that other measures must be used for series operation.
- Structure-borne noise sensors attached to the engine block are inexpensive and provide indirect information about the combustion taking place inside the engine.
- knocking burns can be detected via sound peaks.
- the advantages of the use of structure-borne noise instead of the pressure are paid for with a more complicated and error-prone evaluation, because other effects can be noticeable in structure-borne noise.
- neural networks are difficult to use and do not always lead to reproducible results since many parameters (network structure, transfer functions) must be specified a priori.
- the weights of the network must be optimized numerically numerically, often only suboptima are found.
- HMMs are an alternative approach.
- the temporal and the spectral variability of the signals in the form of a stochastic automaton will be described on the basis of a given example or training data set.
- the actual structure-borne sound signals are converted into time sequences of spectral vectors using STFT ⁇ "Short Time Fourier Transforms.”
- the temporal pattern of the spectral vectors-the feature sequences-can be modeled by an HMM.
- HMMs can only be used to a limited extent for knock detection, since HMMs can only model relatively short sequences, preferably short, non-stochastic sequences, due to the averaging properties of the states. Furthermore, they have similar disadvantages as neural networks.
- the object of the invention is a method and a device for automatic pattern recognition in a sequence of electronic data by means of electronic data processing in specify a data processing system with which a reliable detection of patterns in the sequence of electronic data can be executed in a simplified manner.
- the object is achieved by a method for automatic pattern recognition according to independent claim 1 and an apparatus for automatic pattern recognition according to independent claim 5.
- the invention encompasses the idea of a method for automatic pattern recognition in a sequence of electronic data by means of electronic data processing in a data processing system, in which automatic electronic pattern recognition in a sequence of electronic data by means of electronic data processing in a data processing system, in an analysis of the sequence of electronic data is compared with parameterized model data representing at least one pattern sequence and in which the at least one pattern sequence is detected when it is determined in the analysis that model data comprised by the parameterized model data associated with the at least one pattern sequence has a similarity threshold Similarity measure occur, wherein in forming the parameterized model data training data by means of a dynamic-time warping method to a set of feature vectors of equal length and with a the same information content as the training data are processed, from which the parameterized model data are derived.
- an apparatus for automatic pattern recognition in a sequence of electronic data by electronic data processing comprising a data processing system comprising: pattern recognition means configured to, in an analysis, sequence the electronic data with parameterized model data comparing at least one pattern sequence and recognizing the at least one pattern sequence if it is determined in the analysis that model data included in the parameterized model data associated with the at least one pattern sequence has a similarity measure exceeding a similarity threshold, and
- Model data generating means configured to generate the parameterized model data using the training data and thereby the training data by means of a dynamic-time warping method to a set of feature vectors of equal length and to process with the same information content as the training data from which the parameterized model data are derived, and
- Providing means configured to provide electronically evaluable recognition information about recognizing the at least one pattern sequence for output.
- a preferred embodiment of the invention provides that the parameterized model data are derived from the set of feature vectors by parameterizing a feature vector-based classifier.
- a Bayes classifier with kernel window density estimation is used as the feature vector-based classifier.
- An expedient development of the invention provides that the similarity measure for a subsequence of electronic data examined at the time j of the analysis is determined from the sequence of electronic data as follows:
- X j are the elements of the sequence of electronic data, p t , (•) and p e , (•) the z-th elements of a total of N elements of the parametric model data and c and a m are constants to be chosen empirically.
- the sought similarity measure at time y is L (NJ).
- the method may be used in conjunction with various automatic pattern recognition technologies including, but not limited to, machine signal analysis such as engine knock analysis, ECG signal analysis, speech recognition, gene sequence analysis, image analysis, and thermal image data evaluation.
- machine signal analysis such as engine knock analysis, ECG signal analysis, speech recognition, gene sequence analysis, image analysis, and thermal image data evaluation.
- speech recognition for the quality control of machine-forged components belong.
- thermal image data evaluation for the quality control of machine-forged components belong.
- the data to be analyzed and the example and training data in electronic form and corresponding measurement or analysis variables are available.
- Fig. 1 is a schematic representation of a structure of a knock control for an engine
- Fig. 2 shows an example of the data to be processed in the knock control
- Fig. 3 is a schematic representation which describes the relationship between measured structure-borne sound signals and sequentially arranged electronic data.
- the pattern recognition method comprises three sub-aspects that can be considered separately, namely (i) a data set transformation, (ii) a parameter determination of a model, and (iii) the application of the parameterized model for recognizing sequences or sequences in sequentially arranged electronic data in turn can represent a wide variety of information content.
- a transformation of an example or training data set into feature vectors takes place, which makes hidden random variables accessible and direct comparability possible. It is assumed that there are three training or example sequences for the parameter determination:
- the example or training data set each represents electronically evaluable information about one or more patterns of measurable size to be later recognized.
- 5 3 ⁇ a, *, *, b, b, b r c r *, d, d r *, e r *, £, £, *, g, g).
- the star symbols can be replaced without loss of information by the predecessor symbols, as always a back transformation would be possible by the attached binary vectors and there are the feature vectors
- a probability density p (m) can be estimated. This describes the structure and randomness of the data both in time and in amplitude.
- a kernel approach for example a Parzen approach, can be used (Parzen: On estimation of a probability density and mode, Annais of Mathematical Statistics, VoI 33: 1065-1076, 1962):
- n is the number of feature vectors
- d the dimension of the feature vectors
- s (si, ..., S r J ⁇ is a smoothing parameter to be estimated
- m ⁇ (mu, ..., w ⁇ r is the k th feature vector
- Gaussian functions 0 (m - m ,, s) and ⁇ (m - m ⁇ s) are combined with i ⁇ j into a single Gaussian function ⁇ ' ⁇ (m - m', sj) whose similarity is big enough.
- the new parameters occur as a result of the forming process , s' and m'on.
- the resulting model of the distribution is after the
- the vector dimension d can then be reduced in the same way.
- Each of the resulting q Gaussian functions ⁇ ⁇ va. - va k ',%' k ) is a specialist for a subset of the data and consists of a product of scalar Gaussian functions.
- the scalar Gaussian functions thereby model either a local probability density in time or in amplitude, depending on the component of the feature vector m, which consists of a sequence S and a binary distortion vector ⁇ .
- Emission densities and transition densities are merely the factors of the product (9) in recoded form.
- the pararnetrization phase is over. The following part describes how the model can be applied efficiently. The sub-aspect concerning the application of the actual pattern recognition model follows.
- the method works like a digital filter, i.
- a measure is output which gives information about the current similarity. If this similarity measure exceeds a given threshold, then a suitable appearing reaction can take place.
- the evaluation of the sequence S is also possible synchronously to a measurement, since only the current measured value is needed.
- the probability distributions p x ⁇ (-) and />,, (•) result from the relation (10).
- the parameter a m is at least as large to choose, so that applies to all p n (a m ) «0.
- the value L (N, j) is the sought similarity measure at time j, which indicates how closely the currently observed sequence resembles one of the sequences from the parameterization phase. Overall, there are q of these values. The largest of these is relevant and is compared to the detection threshold to signal a detection event when it is exceeded.
- An implementation of L (i, j) in the form of a ring buffer is possible.
- the method described above describes in a general way the proposed process of pattern recognition, as it can be used in various application cases. In the following, application examples for the use of the pattern recognition method will now be described in more detail.
- Fig. 1 shows a schematic representation of a structure of a knock control for a motor.
- a structure-borne sound signal is continuously recorded and digitized by means of an analog-to-digital conversion with a sufficiently high sampling rate.
- the time signal thus becomes a sequence of scalars.
- this sequence is converted by means of an STFT into a sequence of spectral vectors (spectrogram: amplitude spectrum or power density spectrum), which describe the expression of certain frequency components over time.
- the spectral vectors can then be logarithmized and converted into cepstral vectors by means of a discrete cosine transformation.
- This step is not mandatory.
- the vector sequences will be referred to hereinafter as feature vector sequences to abstract from the specific type of preprocessing that is completed. The actual recognition takes place exclusively on the basis of these feature vector sequences as generally explained above.
- example or training data must be recorded with the help of an engine test stand.
- the type of engine to be controlled is placed at different speeds and for each cylinder in the knocking and non-knocking area.
- the cylinder internal pressure is measured. These data are required in order to be able to clearly judge whether a concretely measured structure-borne noise signal corresponds to a knocking or a non-knocking combustion (see Fig. 2).
- the recorded structure-borne noise data are prepared by cutting out all areas in which there is an overpressure in the simultaneously measured pressure signal.
- the knocking strength of each structure-borne sound fragment is determined on the basis of the pressure signal and connected to it (labeled).
- the pressure signals are bandpass filtered and rectified. The remaining maximum amplitude represents a measure of the current strength of knocking.
- a data set of structure-borne sound fragments is available, with which the knock detection can be parameterized. The pressure signals are then no longer needed.
- the first model is for the detection of knocking burns, the second for the detection of non-knocking burns. In this way the task can be reduced to a simple classification problem.
- the starting point for the parameterization are the structure-borne sound fragments cut out of the continuous structure-borne noise signal and labeled with the knocking strength.
- the model for non-knocking burns is parameterized only with those structure-borne sound fragments whose knock strength lies below a previously defined threshold S 1 . Accordingly, the model for the knocking burns is parameterized with the help of clearly knocking structure-borne sound fragments.
- the knock intensity must exceed a threshold ⁇ 2 .
- Both thresholds S 1 and s 2 may be the same. However, it makes sense to choose S 2 slightly larger than S x . Apart from the database used, both models are otherwise completely identical. Likewise, the parameterization phase is not different from each other, so it is sufficient to describe them using a single model.
- the pattern recognition it is more favorable for the pattern recognition not to analyze the structure-borne sound signals directly, but rather feature vector sequences derived therefrom, ie sequences of feature vectors.
- a structure-borne sound fragment thus becomes a feature vector sequence (see Fig. 3). Since the structure-borne sound fragments differ in their length, the feature vector sequences generated by the preprocessing differ in their length. A direct comparison is not possible.
- dealing with the classification problem with classical feature vector-based pattern recognition methods is impossible because they require that a self-contained feature space exists and thus allow implicit estimation of the probability distribution of the example data set.
- model for pattern recognition can be used in the manner explained above. Since two models were generated during the parameterization phase, namely once for knocking and once for non-knocking burns, two of these values exist. Depending on which of these values is greater, either a knocking or a non-knocking combustion is present. If both values are low, there is either no combustion at the moment or the sensor is damaged. The engine control unit thus has the opportunity to detect a failure of the knock detection, which is important in order to avoid damage to the engine.
- the method described allows a continuous search for knocking burns.
- the method like a digital filter, can provide a criterion for the instantaneous knock magnitude at each sampling instant.
- no a priori specifications are required and the determination of the parameters is largely constructive, i. without numerical optimization.
- Some of the applications are based on time signals. In these applications, it is relatively obvious at which point the method of sequencing is usefully employed. can be set. For example, in the signal analysis of ECG signals (ECG electrocardiogram) directly the time signal can be used. It is then a use of the above-described method for automatic pattern recognition in a signal analysis of ECG signals. In this way, sequences in the ECG signals can be determined, which may indicate arrhythmias.
- a model For each command word a model is created.
- the corresponding examples are preprocessed and converted into spectral vector sequences. These are the actual sequences from which characteristic vectors of the same length are then generated in the manner already described (formulas (1) to (4)). With the aid of the described parameterization (formulas (5) to (10)) the models are subsequently generated.
- the relationship (11) then allows the use of the generated models to analyze a continuous audio signal. If the similarity measure for each model constantly calculated to a certain If the time exceeds the predefined threshold, it can be assumed that the continuously examined audio signal currently contained an utterance which was similar to the command words used in the parameterization of the corresponding model. A message of the associated label appears to the user of the system as recognition of his spoken utterance and can be used to trigger certain useful actions.
- the patterns to be searched consist of certain significant code fragments, ie sequences or sequences of bytes describing the behavior of the code.
- variations are often added to certain parts of the code that, while not modifying the actual behavior, result in a changed sequence of bytes.
- NOP machine instructions No Operation
- the procedure for locating malicious program code using the method described above is to describe the byte sequences of different modified versions by a common model and to search for the occurrence of the virus with this. For this, the byte sequences of the formulas (1) to (4) are correspondingly transformed into feature vectors of fixed length. Subsequently, the parameterization of the model takes place. It is then a use of the above-described method for automatic pattern recognition in virus scanning.
- a handwritten text can be interpreted as a sequence or sequence of XY coordinates.
- these sequences can not be directly compared.
- the invention provides a direct way of processing such data.
- the task could be to check the signature or signature of a person, e.g. to authenticate a laptop.
- the necessary hardware, a touchpad and a computer for the evaluation are already included in the devices.
- Each sequence begins when a touch is registered on the touchpad and ends when it has not been touched for a while.
- the first coordinate of the sequence may be subtracted from all remaining coordinates of the sequence. This will ensure that each coordinate sequence starts at the origin (0,0).
- time signals are often used which can be interpreted directly as sequences, namely current or voltage characteristics.
- Other sensor data in which interference by transfer functions takes place can be examined in the form of spectrograms (see knock detection above).
- sequence recognition can be sensibly used.
- it is typical that these are almost always detail problems, for example part of a controller, part of process monitoring or the like. It is then a use of the above-described method for automatic pattern recognition in the control or process monitoring of a machine or plant, wherein the sequence of electronic data represents data acquired for the control or the process monitoring, whereby previously associated sample or training data is acquired were.
- Another application of the pattern recognition method is the evaluation of thermal image data for quality control of machined forged components.
- Forged components occasionally show cracks. Visually, the cracks are usually not easy to recognize. However, the respective cooling behavior deviates from areas with cracks and areas without cracks.
- images of the forged components are recorded by means of a thermal imaging camera for a short time.
- the cooling of a component corresponds to a change in a gray value image G (x, y, t) formed by the thermal imaging camera over a time t. Since the position of the component with respect to the thermal imaging camera does not change during the recording, the image coordinates x and y (pixels) are assigned to a respective area of the component surface.
- the temporal behavior of the gray value can be approximately described here by means of a decaying exponential function:
- the parameter I (x, y) can preferably be estimated by means of linear regression. Further parameters describing the cooling are possible.
- each column Sp (x) (V (x, l), V (x, 2), V (x, 3), ...) of the secondary image V (x, y) as a vector interpret.
- the task of finding the position of the component and the comparison with a reference is thus reduced to a sequence detection problem which can be solved with the pattern recognition method according to the invention.
- the reference image (reference) is formed for example by means of the method according to the invention from several example sequences of defect-free components
- a method for automatic pattern recognition is described above, which can be used in a variety of applications by analyzing corresponding electronic data, which comprise an information associated with the respective application, in the manner explained above.
- the starting point of the method here is first the generation of a set of feature vectors of equal length or dimension from training or example data by means of a dynamic-time warping method.
- feature vectors are generated, which can then be examined in principle with the aid of any classifiers for pattern recognition.
- a neural network eg a multilayer perceptron
- bis- hop Neural networks for Pattern Recognition, Clarendon Press, Oxford, 1995.
- the generation of the set of feature vectors constitutes an independent aspect of the invention, which develops its advantages independently of the subsequent choice of the classifier and thus in connection with various classifiers.
- the described method for automatic pattern recognition can advantageously be used in particular in connection with the following applications: machine speech recognition, handwriting recognition, gene sequence analysis, search for malicious program code (virus scanner), medical technology applications such as cardiac pacemakers or electrocardiograms and mechanical diagnostic applications such as knock detection.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Biology (AREA)
- Software Systems (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Databases & Information Systems (AREA)
- Health & Medical Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Multimedia (AREA)
- Computing Systems (AREA)
- Medical Informatics (AREA)
- General Health & Medical Sciences (AREA)
- Image Processing (AREA)
- Complex Calculations (AREA)
- Image Analysis (AREA)
Abstract
L'invention concerne un procédé de reconnaissance automatique de motifs dans une séquence de données électroniques au moyen d'un traitement électronique de données dans un système de traitement de données, dans lequel, au cours d'une analyse, la séquence de données électroniques est comparée à des données paramétrisées de modèle qui représentent au moins une séquence de motifs et dans lequel la ou les séquences de motifs sont reconnues si l'analyse détermine que des données de modèle constituées des données de modèle paramétrisées qui sont associées à la ou aux séquences de motifs surviennent à un niveau de similitude qui dépasse un seuil de niveau de similitude, tandis que lors de la formation des données de modèle paramétrisées, des données d'apprentissage sont traitées au moyen d'un procédé de gauchissement temporel dynamique pour former un ensemble de vecteurs caractéristiques de même longueur, dont le contenu informatif est le même que celui des données d'apprentissage et à partir desquels les données de modèle paramétrisées sont déduites. L'invention concerne en outre un dispositif de reconnaissance automatique de motifs dans une séquence de données électroniques au moyen d'un traitement électronique de données doté d'un système de traitement de données.
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
DE102007036277A DE102007036277A1 (de) | 2007-07-31 | 2007-07-31 | Verfahren und Vorrichtung zur automatischen Mustererkennung |
PCT/DE2008/001256 WO2009015655A2 (fr) | 2007-07-31 | 2008-07-31 | Procédé et dispositif de reconnaissance automatique de motifs |
Publications (1)
Publication Number | Publication Date |
---|---|
EP2174267A2 true EP2174267A2 (fr) | 2010-04-14 |
Family
ID=40175840
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
EP08801094A Withdrawn EP2174267A2 (fr) | 2007-07-31 | 2008-07-31 | Procédé et dispositif de reconnaissance automatique de motifs |
Country Status (4)
Country | Link |
---|---|
US (1) | US20100217572A1 (fr) |
EP (1) | EP2174267A2 (fr) |
DE (1) | DE102007036277A1 (fr) |
WO (1) | WO2009015655A2 (fr) |
Families Citing this family (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
TWI467568B (zh) * | 2007-07-13 | 2015-01-01 | Dolby Lab Licensing Corp | 使用位準時變評估機率密度之時變音訊信號位準 |
DE102015204208B4 (de) * | 2015-03-10 | 2024-09-26 | Bayerische Motoren Werke Aktiengesellschaft | Verfahren und Steuereinheit zur Überwachung einer Kommunikationsverbindung |
CN110634050B (zh) * | 2019-09-06 | 2023-04-07 | 北京无限光场科技有限公司 | 一种鉴别房源类型的方法、装置、电子设备及存储介质 |
US11281917B2 (en) * | 2019-10-31 | 2022-03-22 | Aptiv Technologies Limited | Multi-domain neighborhood embedding and weighting of point cloud data |
CN111694331B (zh) * | 2020-05-11 | 2021-11-02 | 杭州睿疆科技有限公司 | 生产工艺参数调整的系统、方法和计算机设备 |
Family Cites Families (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPH02263275A (ja) * | 1989-04-03 | 1990-10-26 | Kiyadeitsukusu:Kk | 手書き文字の登録パターン作成方式 |
DE69425166T2 (de) | 1993-02-26 | 2001-03-15 | Canon K.K., Tokio/Tokyo | Verfahren und Gerät zur Mustererkennung |
EP0909426B1 (fr) | 1996-07-05 | 2001-03-07 | Osmetech PLC | Reconnaissance de formes a l'aide d'un reseau neuronal |
AU713371B2 (en) | 1996-07-29 | 1999-12-02 | British Telecommunications Public Limited Company | Pattern recognition |
DE19650541C2 (de) * | 1996-12-05 | 1999-05-12 | Siemens Ag | Verfahren zur Ermittlung eines ersten Referenzschriftzugs anhand mehrerer Musterschriftzüge |
DE19741884C2 (de) * | 1997-09-23 | 2000-12-21 | Daimler Chrysler Ag | Verfahren zur Bestimmung relevanter Größen, die den Zylinderdruck in den Zylindern einer Brennkraftmaschine repräsentieren |
DE50107442D1 (de) | 2000-08-11 | 2005-10-20 | Bosch Gmbh Robert | Klopferkennung bei brennkraftmaschinen mit modifizierung bei änderung einer filtercharakteristik oder zylinderindividueller änderung |
DE10043498A1 (de) * | 2000-09-01 | 2002-03-14 | Bosch Gmbh Robert | Verfahren zur klopferkennung bei Brennkraftmaschinen |
KR100580618B1 (ko) * | 2002-01-23 | 2006-05-16 | 삼성전자주식회사 | 생리 신호의 단시간 모니터링을 통한 사용자 정서 인식장치 및 방법 |
DE10300204A1 (de) | 2003-01-08 | 2004-07-22 | Robert Bosch Gmbh | Verfahren und Vorrichtung zur Klopferkennung |
US8346482B2 (en) * | 2003-08-22 | 2013-01-01 | Fernandez Dennis S | Integrated biosensor and simulation system for diagnosis and therapy |
DE10352860B4 (de) | 2003-11-10 | 2013-10-31 | Iav Gmbh Ingenieurgesellschaft Auto Und Verkehr | Verfahren zur Auswertung miteinander korrelierender Messdaten |
US7223234B2 (en) * | 2004-07-10 | 2007-05-29 | Monitrix, Inc. | Apparatus for determining association variables |
US7519461B2 (en) | 2005-11-02 | 2009-04-14 | Lear Corporation | Discriminate input system for decision algorithm |
-
2007
- 2007-07-31 DE DE102007036277A patent/DE102007036277A1/de not_active Ceased
-
2008
- 2008-07-31 WO PCT/DE2008/001256 patent/WO2009015655A2/fr active Application Filing
- 2008-07-31 EP EP08801094A patent/EP2174267A2/fr not_active Withdrawn
- 2008-07-31 US US12/671,248 patent/US20100217572A1/en not_active Abandoned
Non-Patent Citations (1)
Title |
---|
See references of WO2009015655A2 * |
Also Published As
Publication number | Publication date |
---|---|
WO2009015655A3 (fr) | 2009-03-26 |
DE102007036277A1 (de) | 2009-02-05 |
WO2009015655A2 (fr) | 2009-02-05 |
US20100217572A1 (en) | 2010-08-26 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
DE102018200724A1 (de) | Verfahren und Vorrichtung zum Verbessern der Robustheit gegen "Adversarial Examples" | |
DE202017102235U1 (de) | Trainingssystem | |
EP0780002B1 (fr) | Procede et appareil pour reconstruir de structures constituees de lignes sous forme d'une trame | |
EP1193688A2 (fr) | Procédé pour déterminer un espace propre pour la présentation d'une pluralité de locuteurs d'apprentissage | |
DE102017220307A1 (de) | Vorrichtung und Verfahren zum Erkennen von Verkehrszeichen | |
WO2009015655A2 (fr) | Procédé et dispositif de reconnaissance automatique de motifs | |
EP2172820A1 (fr) | Procédé et système destinés à l'analyse automatique de données de processus | |
DE69805280T2 (de) | Gerät und verfahren zur mustererkennung. | |
EP1193687A2 (fr) | Adaptation au locuteur pour la reconnaissance de la parole | |
DE102019211672A1 (de) | Trainingsverfahren für ein künstliches neuronales Netzwerk | |
DE69132148T2 (de) | Vorrichtung zur Verarbeitung eines Signals | |
EP3786853A1 (fr) | Compression d'un réseau de neurones profond | |
DE102023207534A1 (de) | System und Verfahren zur universellen Bereinigung von Eingangsstörungen mit entrauschten Diffusionsmodellen | |
EP3857455A1 (fr) | Système d'apprentissage automatique ainsi que procédé, programme informatique et dispositif pour créer le système d'apprentissage automatique | |
DE69715343T2 (de) | Abstandsberechung zur Verwendung in einem Spracherkenner | |
WO2001086585A1 (fr) | Procede et dispositif de determination d'un objet dans une image | |
DE102021204040A1 (de) | Verfahren, Vorrichtung und Computerprogramm zur Erstellung von Trainingsdaten im Fahrzeug | |
EP3701428B1 (fr) | Procédé et dispositif destinés à améliorer la robustesse d'un système d'apprentissage par machine | |
WO2006134011A1 (fr) | Procede de traitement de donnees numeriques assiste par ordinateur | |
EP1400951B1 (fr) | Méthode informatisée pour la reconnaissance de la parole, système de reconnaissance de parole et un système de contrôle pour un système technique et système de télécommunication | |
EP3973458A1 (fr) | Procédé pour faire fonctionner un réseau neuronal profond | |
EP1889754B1 (fr) | Procédé et dispositif destinés à commander des moyens de protection des personnes et produit de programme informatique | |
WO2002017127A2 (fr) | Procede et dispositif d'analyse de correlation de sequences de donnees | |
EP2259208A2 (fr) | Procédé d'établissement de champs d'orientation des lignes papillaires d'une empreinte de doigts | |
DE10144661C2 (de) | Vorrichtung und Verfahren zur Zuordnung der NMR-Signale von Polypeptiden |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PUAI | Public reference made under article 153(3) epc to a published international application that has entered the european phase |
Free format text: ORIGINAL CODE: 0009012 |
|
17P | Request for examination filed |
Effective date: 20100128 |
|
AK | Designated contracting states |
Kind code of ref document: A2 Designated state(s): AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HR HU IE IS IT LI LT LU LV MC MT NL NO PL PT RO SE SI SK TR |
|
AX | Request for extension of the european patent |
Extension state: AL BA MK RS |
|
17Q | First examination report despatched |
Effective date: 20100823 |
|
DAX | Request for extension of the european patent (deleted) | ||
STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: THE APPLICATION IS DEEMED TO BE WITHDRAWN |
|
18D | Application deemed to be withdrawn |
Effective date: 20140201 |