PatentDe  


Dokumentenidentifikation EP0602932 19.04.2001
EP-Veröffentlichungsnummer 0602932
Titel Verbessertes System zur Mustererkennung für Sonar und andere Anwendungen
Anmelder Raytheon Co., El Segundo, Calif., US
Erfinder Holmberg, Bart A., Bellevue, Washington 98005, US
Vertreter Kuhnen & Wacker Patentanwaltsgesellschaft mbH, 85354 Freising
DE-Aktenzeichen 69330021
Vertragsstaaten DE, FR, GB
Sprache des Dokument EN
EP-Anmeldetag 14.12.1993
EP-Aktenzeichen 933100661
EP-Offenlegungsdatum 22.06.1994
EP date of grant 14.03.2001
Veröffentlichungstag im Patentblatt 19.04.2001
IPC-Hauptklasse G06K 9/62

Beschreibung[en]
BACKGROUND OF THE INVENTION Field of the Invention:

The present invention relates to pattern recognition systems. More specifically, the present invention relates to techniques for building classifiers for pattern recognition systems.

Description of the Related Art:

In many applications, there is a need for a capability to recognize patterns in samples of data. In long range sonar applications, for example, there is a need to recognize targets in a background of clutter. As pattern recognition is difficult to achieve with conventional sequential processors, neural network (net) processors are typically employed for this application. Artificial neural networks utilize an array of relatively simple processors in a parallel fashion much like the brain. The processors with associated nonlinearities and interconnection weights provide a classifier which can implement a classification algorithm. The classifier determines which class of data a sample of data was most likely to have come from. For this purpose, the classifier must be trained. Training involves the application of known data, to an algorithm which modifies the classifier weights.

For many applications, classifier training is critical and expensive. In the long range active sonar application, for example, classifier training typically involves the use the tracking of a known target, e.g., a submarine, which transmits position to the tracking system. This target position information is used to process the data received from the sonar receiver and thereby extract features representative of the clutter. These extracted features are used to train the classifier. The adaptation or training of the classifier requires an operator to overlay a representation of the target on a displayed received image and manually classify data samples as having come from the target class or the clutter class. With thousands of data points for a single frame of data, this process is typically time intensive and quite expensive. In any event, data thus obtained is then used to train classifiers in the sonar data processing systems of numerous other submarines or other towed array systems in the fleet.

While this expensive technique may be adequate in applications where the statistics of the unknown class are stationary, it is severely limited, if not ineffective, where the statistics are extremely dynamic such as in long range sonar applications.

Hence, the conventional approach to the training of classifiers for pattern recognition systems is limited by the need for statistics on both a known class and an unknown class and the need for a stationary statistical model of the data for the unknown class.

Thus, there is a need in the art for an automated technique for building classifiers for nonstationary data classes. There is a further need in the art for a classifier training technique which is not limited by the need for more than one class of data.

SUMMARY OF THE INVENTION

The need in the art is addressed by the present invention which provides an improved pattern recognition system and method as claimed in the appendant claims. The invention operates on a plurality of feature vectors from a single class of data samples. The inventive system estimates a pruning radius for the feature vectors in the single class of data samples and generates a replacement class therefrom based on the estimated pruning radius. This pruning radius is used to train a classifier which in turn facilitates the recognition of a data pattern in raw data. The pruning radius is adapted based on current results from the classifier.

The invention satisfies the need in the art by providing an automated technique for training classifiers for nonstationary environments which is not limited by the need for more than two classes of data.

BRIEF DESCRIPTION OF THE DRAWINGS

Fig. 1 is a diagram showing two submarines engaging in a conventional classifier training exercise.

Fig. 2 is a block diagram of a sonar system equipped with a conventional pattern recognition system.

Fig. 3 is a diagram of input feature data useful in explaining the operation of a conventional classifier.

Fig. 4 is a diagram representative of boundary formation of a conventional classifier.

Fig. 5 is a block diagram of a sonar system having the improved pattern recognition system of the present invention.

Fig. 6a is a block diagram of the classifier training system of the present invention in a training mode of operation to find classifier weights.

Fig. 6b is a block diagram of the classifier training system of the present invention in acting as an adaptive classifier in a long range active application.

Fig. 7a is a diagram of a two dimensional feature space useful in the disclosure of the present teachings.

Fig. 7b is a diagram of a two dimensional feature space where the pruning radius and average nearest neighbor distance are equal because the variance is zero.

Fig. 7c is a diagram of a two dimensional feature space where the pruning radius is equal to the average nearest neighbor distance plus a standard deviation.

Fig. 8 is a diagram of input feature data useful in explaining the advantageous operation of the present invention.

Fig. 9 is a diagram of feature data after a first number of random presentations of input data to the classifier of the present invention.

Fig. 10 is a diagram of feature data after a second number of random presentations of input data to the classifier of the present invention.

DESCRIPTION OF THE INVENTION

Illustrative embodiments and exemplary applications will now be described with reference to the accompanying drawings to disclose the advantageous teachings of the present invention.

The conventional long range sonar system requires use of a pattern recognition system to discriminate target data from background data. Pattern recognition is facilitated by a classifier which separates the data according to predetermined features along a boundary. For this purpose, the classifier must first be trained. The conventional classifier training technique involves the receipt of data from an unknown class (background) and the receipt of data from a known class (target data). Conventionally, data from the known class is generated by the transmission of position data from a known target to the tracking system while simultaneously receiving data on the unknown class. In a long range active sonar application, for example, this would involve the use of two submarines with one tracking the other.

Fig. 1 is a diagram showing two submarines engaging in a conventional classifier training exercise. The tracking submarine 1 emits a sonar signal which is reflected back to the tracking submarine by the tracked submarine 2. The tracked submarine 2 acquires its position from a Global Positioning System (GPS) satellite 3, for example, and transmits its position to the tracking submarine 1 directly or via a relay link not shown. The tracking submarine 1 receives the sonar return along with returns from numerous other locations including the ocean surface 4 and the ocean floor 5. These returns represent background clutter to the sonar receiver of the tracking submarine.

Fig. 2 is a block diagram of the sonar system 10' of the tracking submarine 1 equipped with a conventional pattern recognition system. The system 10' includes a sonar transmitter 11' and a sonar receiver 12'. The receiver 12' includes a phased array of sonar receiving elements 14' and a beamforming network 16' as is common in the art. The output of the beamforming network 16' is a signal corresponding to the generated beam. The signal is processed by a signal processor 18' to provide raw feature data. The signal processor 18' typically includes a bank of matched filters, detection and normalization circuits. A feature extractor 20' equipped with a classifier (not shown) operates on the raw feature data, identifies features (feature vectors) in the data samples and classifies the features in accordance with a feature boundary or surface in a multidimensional plane. The classified features are input to a discriminator 22' which endeavors to recognize patterns therein. The feature surface is provided initially in accordance with predetermined parameters. Thereafter, the feature surface is updated by the conventional classifier training system 24'. The conventional classifier training system 24' includes a data processor 26' which receives pattern data from the discriminator 22' and position data from the tracked submarine 2 (represented in the figure as a "KNOWN DATA CLASS SOURCE" 28') and displays representations of both on a display 30'.

In accordance with conventional teachings, an operator trains the classifier by examining the display and manually classifying data in and around the known class space. The classifier, thus trained, is then used to train the classifiers of other submarines.

Fig. 3 is a diagram of input feature data useful in explaining the operation of a conventional classifier and a classifier. The two classes are '*' and 'o' where 'o' represents the class of known data. Both are distributed evenly uniformly randomly within the optimal and as yet unrecognized circle boundaries shown. There are 800 sample vectors from each class. The larger spheres 34, 36 and 38 represent the feature vectors from the known class and the smaller spheres 40, 42, 44 and 46 represent feature vectors from the unknown class. Given both classes, to facilitate pattern recognition, a boundary should be found that separates the classes.

Fig. 4 is a diagram representative of boundary formation of the conventional classifier after 560,000 simulated presentations of the input feature data of Fig. 3 thereto. The samples below the boundary 48 were classified as being in a first class and samples above the boundary 48 were classified as being in a second class. It is evident that, in the simulation, the illustrative conventional classifier misclassified feature vectors in boundary spheres 40 and 42. In this example, this amounts to a 50% error classification rate.

In addition to a susceptibility for classification errors, this conventional classification training process is quite expensive inasmuch as it requires that data be generated from a known class, i.e., by the tasking and tracking of a second submarine, and the manual classification of feature vectors. Accordingly, it is particularly unfortunate that the data is typically nonstationary. As a result, the classifiers trained in accordance with the conventional technique are often grossly inaccurate after the passage of a relatively short period of time. Accordingly, there has been a need in the art for an automated technique for training classifiers for nonstationary data classes which is not limited by the need for more than one class of data.

This need is addressed by the present invention which provides an improved pattern recognition system which operates on a plurality of feature vectors from a single class of data samples. The inventive system estimates a pruning radius for the feature vectors in the single class of data samples and generates a replacement class therefrom based on the estimated pruning radius. This pruning radius is used to train a classifier which in turn facilitates the recognition of a data pattern in raw data. The pruning radius is adapted based on current results from the classifier.

Fig. 5 is a block diagram of a sonar system having the improved pattern recognition system of the present invention. As with the conventional system 10', the inventive system 10 includes a sonar transmitter 11, a sonar receiver 12, a signal processor 18, a feature extractor 20 and a discriminator 22. The inventive system differs, however, from the conventional system in that the classifier trainer 24' is eliminated along with the need for a source of a second class of known data and the need for operator intervention. The classifier trainer 24' of the conventional pattern recognition system 10' is eliminated by the use of the novel classifier training system of the present invention.

The inventive classifier consists of a binary classifier training method.

Fig. 6a is a block diagram of the classifier training system of the present invention in a training mode of operation to find classifier weights.

The novel classifier training system 100 includes a pruning radius estimator 110, a replacement class generator 120, a backpropagation trainer 130 and a classifier 140. Fig. 6a shows how the inventive classifier trains on a single class of features to find the classifier weights.

Fig. 6b is a block diagram of the classifier training system of the present invention in acting as an adaptive classifier in a long range active application.

The method of operation of the classifier training system 100 is comprised of two main parts. The first part involves the selection of which of the two classes is easiest to obtain. The other class will then be represented by a replacement class. The replacement class will consist of random feature vectors which are drawn from a uniform distribution. Each element of the replacement class will consist of random feature vectors which are drawn from a uniform distribution. Each element of the replacement class vector is then a random variable, uniformly distributed over the range of possible values of that feature (e.g., feature n is the output of a photodetector which has maximum output magnitude 0.1 volt).

The second part involves modification of the completely random replacement class with the information contained in the known class. The modification consists of subtracting the space which contains the known class from the replacement class. The method for accomplishing this involves deleting all vectors drawn from the replacement class which are within a hypersphere of radius H of an known class sample.

The statistics of the process of deleting the replacement vectors falling in the known class space from the uniform replacement class are derived for this problem assuming uniform, independent known class statistics. This is done in terms of the hypersphere radius H. For real world applications, the uniform, independent assumption on the known class distribution may not be very accurate. Since the statistics have not been derived for a more realistic model (jointly Gaussian for example), a heuristic method for determining the distance H is as follows.

The pruning radius H is illustrated in Fig. 7a as the average distance between samples in the known class. The pruning radius may be determined in accordance with either of two methods. The first method is a statistical technique which finds the radius H in terms of the probability of a replacement class sample occurring in the known class space. This technique makes strong assumptions about the type of class distributions and knowledge of it. The second technique makes no assumption about the underlying distributions but does not provide a nice relationship between the pruning radius and probability of a replacement class sample falling in the known class space. It does however hold some heuristic appeal. It should also be noted that there is a tradeoff in the radius H. If it is made too small, replacement class samples will remain in the known class space. If H is too large, the boundary that the classifier finds will not fit tightly about the known class space for the uniform case or for Gaussian like distributions, the boundary will favor the known class space.

The first method for determining the pruning radius H is as follows. First, the following assumptions are made: 1) the volume of the known class distribution is known or can be estimated and 2) distribution of the known class is uniform over its volume.

Fig. 7a is a diagram of a two dimensional feature space useful in the disclosure of the present teachings. The distributions A, B, and C are defined as having volumes VA, VB, and VC, respectively. The point density λA, λB and λC are defined as equal to N/V where N is the number of sample points and V is the volume. Event C is defined as the placement of a hypersphere of volume VC arbitrarily within the space contained by VB and no sample points from distribution B fall within VC. From the uniform distribution assumption and the Poisson pdf The probability of event C is given as follows: Pr{Event C} = eBVC

Event L is defined as the placement of M hyperspheres of volume VC randomly inside of the space defining VB. None of the spheres are empty. Performing a Bernoulli trial M times yields:

With simplification, this becomes: Pr{Event L} = (1-eBVC)M

The parameter M is the expected number of replacement distribution sample vectors falling within VB. Thus, M = λAVB.

Defining Pr{Event L} = α, the volume of the hypersphere may be expressed in terms of α. VC = -log(1-elog(α)/(λAVB) / (λB)

Thus, the pruning radius of the hypersphere for the two dimensional case is: H = (-log(1-elog(α)/(λAVB))/λBπ)1/2

This suggests that a should be set to some high probability (i.e., 90%) to insure that we prune all the samples of the replacement distribution from the volume of distribution A.

If either assumption for the first method is not met (i.e., the second assumption is most likely to be invalid for real world data), then another method for determining the pruning radius must be employed. The second method for determining the pruning radius is makes no assumptions about the underlying distributions and provides no direct relation between the pruning radius H and the probability of missing a replacement class sample within a known class distribution. Hence, the second method of determining the pruning radius is as follows.

First, let Vk represent the kth feature vector of known distribution. There are N feature vectors in the known sample distribution. The nearest neighbor distance: dk = min[Vk - VL] for L = 1 to M, except for L = k. The sample average of nearest neighbor distance:

Sample variance of the nearest neighbor distance:

For illustration, assume H=D and the feature space is two dimensional. Assume further that the known distribution samples are spaced at grid points as shown in Figs. 7b and 7c. Fig. 7b is a diagram of a two dimensional feature space where the pruning radius and average nearest neighbor distance are equal because the variance σD is zero. In this case, there is no area left uncovered by the circles placed around the known sample points. The circles actually drawn are only illustrative. If every circle were drawn about every point in the mesh all the points would be overlapped by at least 2 circles. For a random distribution of known sample points, the variance will be some positive number. Conceptually, the higher the variance, the larger the pruning radius should be. To accomplish this, the pruning radius H may be set equal to D+σD as depicted in Fig. 7c. Fig. 7c is a diagram of a two dimensional feature space where the pruning radius is equal to the average nearest neighbor distance plus D the standard deviation σD.

Returning to Fig. 6a, after the pruning radius is determined, the replacement class generator 120 uses the pruning radius to generate a replacement class. The replacement class is uniformly distributed over the feature space with the known class space removed. The method for generating the uniform replacement class and subtracting out the known class space is as follows. First, a uniform N-dimensional replacement class is generated. This can be accomplished with a uniform random number generator. The random number generator would provide independent samples for each feature distributed over the possible range of each feature. Once this is accomplished, those replacement class sample vectors falling in the known class space must be removed. This is accomplished by deleting all samples from the replacement class that fall within the pruning radius H for each sample vector in the known class.

In the illustrative embodiment, the pruning radius estimator 110 and the replacement class generator 120 may be implemented in software. The following is a MATLAB M-file (code) for finding the pruning radius and the replacement class:

ILLUSTRATIVE REPLACEMENT CLASS AND PRUNING RADIUS CODE

After the two classes (the known class and the replacement class) are found, a classifier must be used to find a boundary between the two distributions. The only requirement on the classifier is that it be sufficiently sophisticated to be able to enclose the known distribution.

An illustrative implementation of the artificial neural network classifier 140 of the present invention is as follows. The classifier is implemented in software using conventional techniques. The classifier 140 includes an input layer, which does no processing, and three layers of nodes, the final layer of nodes is referred to as the output layer. There are two inputs (one for each feature), then five nodes in the next layer, ten in the next and two nodes in the output layer. The system equations are: kU = kwka    and (k+1)a=f(kU)   1≤k≤M where:

  • k w is (1+k)N by kN weight matrix where iN is the number of nodes at layer i; so 1N is the number of input features and MN is the number of outputs;
  • iU is the output column vector of layer i; and
  • ia is the activation value column vector which is the input to layer i; so ia is the input feature vector and 1+Ma is the output of the network.

Equation [10] does not provide a specific function. In the illustrative embodiment, this function is the Sigmoid function.

The classifier 140 is trained by presenting a feature vector at the input, propagating the input through the network and comparing it to a known output. The error (the difference between the known output and the propagated input is then used to modify the network weights in a manner which lessens the error. This function is performed in the illustrative embodiment by the backpropagation trainer 130 of Figs. 6a and 6b. Those skilled in the art will appreciate that other classifier training schemes without departing from the scope of the present teachings. The backpropagation trainer 130 of the illustrative system may be implemented in software to perform the following operations. Iw(n+1) - kw(n) - ηδE/δkw E = ((M+1)a - t)2 where

  • η where is a learning coefficient and
  • t is a target value vector.

Implementation: kw(N+1) = kw(n) - kΔw(n) kΔw(n) = ηβ.(k+1)a(n) (Some implementations include a "momentum" term which add αΔw(n-1) to equation [14]. β = (k+1)a(n).(1-(k+1)a(n)).kϕ Mϕ = ((M+1)a(n) - t(n)) i-1ϕ = iϕiw' where 2≤i≤M-1.

Fig. 6b, depicts the operation of the classifier training system of the present invention in a long range active sonar processing environment as an adaptive classifier. In this mode, the class A sample feature vector input shown in Fig. 6a is replaced by raw sample feature vectors from the feature extractor 20 of FIG. 5. The raw sample feature vectors are used to adapt the pruning radius. Thus, the pruning radius estimator tightens a statistical mesh about the known class as it learns more about the environment. Thereafter, the replacement class generator defines a boundary between the known and unknown classes with the replacement class.

The classifier weights are trained on data from time interval [t-Δτ, t] and decisions are made about raw data sample vectors at time intervals [>t]. The choice of Δτ is driven by two conflicting requirements. It must be smaller than the stationary time of the random process from which the feature vectors are samples and it must be large enough to provide a good estimate of the clutter feature space. The only assumption concerning the input data is that the ratio of clutter to target feature samples is large enough not to bias the estimate of the clutter space. This is a good assumption for long range active sonar.

Figs. 8 - 10 illustrate the advantageous operation of the present invention.

Fig. 8 is a diagram of input feature data useful in explaining the advantageous operation of the present invention. The figure corresponds to Fig. 3 in that it depicts a target in a background. Two dimensional input sample vectors are depicted for both the known class 'o', and the replacement class '*'. Note that unlike Fig. 3, data is presented with respect to one known class only, not two as required by classifier training systems of conventional teachings. In Fig. 8, the known class is evenly distributed within the circle boundaries.

Figs. 9 and 10 are diagrams of feature data after 160,000 iterations and 560,000 iterations, respectively. That is, after 160,000 and 560,000 random presentations of input data to the system of the present invention with M = 4, η = 0.9, and α = 0.1 and initial values for the layer weights being outcomes of a pseudo-random number generator evenly distributed between +/- 0.1. The fine line is the decision boundary produced after the iterations and the bold line is the optimal boundary. The replacement class was generated from knowledge of the 'o' class samples used the second method for determining the pruning radius described above. These simulation results show that without any knowledge of the target distribution, a boundary is learned that classifies with near perfect accuracy.

Thus, it can be seen that the present invention provides an improved pattern recognition system which operates on a plurality of feature vectors from a single class of data samples. The inventive system estimates a pruning radius for the feature vectors in the single class of data samples and generates a replacement class therefrom based on the estimated pruning radius. This pruning radius is used to train a classifier which in turn facilitates the recognition of a data pattern in raw data. The pruning radius is adapted based on current results from the classifier.

Thus, the present invention has been described herein with reference to a particular embodiment for a particular application. Those having ordinary skill in the art and access to the present teachings will recognize additional modifications applications and embodiments within the scope thereof. For example, the invention is not limited to sonar applications. Nor is the invention limited to a particular classifier or a particular classifier training technique. In addition, the invention is not limited to the technique illustrated for defining a pruning radius or a replacement class. Other systems and techniques for defining these parameters may be used without departing from the scope of the invention.

It is therefore intended by the appended claims to cover any and all such applications, modifications and embodiments within the scope of the present invention.


Anspruch[de]
  1. Verfahren zum Trainieren einer Klassifizierungseinrichtung (100), welches folgende Schritte enthält:
    • Erzeugen einer Mehrzahl von Merkmalsvektoren aus einer einzelnen Klasse von beispielsweisen, nichtstationären Datenproben;
    • Abschätzen (110) eines Umgrenzungsradius, basierend auf dem Abstand zwischen Proben in der genannten einzelnen Klasse für die genannten Merkmalsvektoren in dieser einzelnen Klasse nichtstationärer Datenproben, wobei der Umgrenzungsradius der radiale Abstand von jedem Merkmalsvektor ist, jenseits welchem ein unbekannter Merkmalsvektor, der für die Klassifizierung dargeboten wird, als ausserhalb der genannten einzelnen Klasse eingeordnet wird, wie sie durch den genannten bekannten Merkmalsvektor definiert wird;
    • Erzeugen (120) einer Austauschklasse von Datenproben aus der genannten einzelnen Klasse von Datenproben, basierend auf dem Umgrenzungsradius unter Erzeugung einer gleichförmigen Verteilung von statistischen Austausch- Vektorelementen, welche einen Gruppenraum für die genannte Austauschklasse von Datenproben einnehmen, wobei die Vektorelemente aus einer gleichförmigen Verteilung ausgewählt sind, sowie unter Löschung aller Vektorelemente innerhalb einer Hypersphäre des genannten Umgrenzungsradius eines Merkmalsvektors aus der genannten einzelnen Klasse zur Erzeugung der genannten Austauschklasse; und
    • Trainieren (130) der genannten Klassifizierungseinrichtung auf der Basis der beiden Klassen, welche durch den Umgrenzungsradius definiert sind, wobei das Trainieren die Bestimmung einer Mehrzahl von Klassifizierungsgewichtungen umfaßt.
  2. Mustererkennungssystem (10) mit einer Klassifizierungseinrichtung, welche durch folgendes Verfahren trainiert wird:
    • Erzeugen einer Mehrzahl von Merkmalsvektoren aus einer einzelnen Klasse von beispielsweisen, nichtstationären Datenproben;
    • Abschätzen (110) eines Umgrenzungsradius, basierend auf dem Abstand zwischen Proben in der genannten einzelnen Klasse für die genannten Merkmalsvektoren in dieser einzelnen Klasse nichtstationärer Datenproben, wobei der Umgrenzungsradius der radiale Abstand von jedem Merkmalsvektor ist, jenseits welchem ein unbekannter Merkmalsvektor, der für die Klassifizierung dargeboten wird, als außerhalb der genannten einzelnen Klasse eingeordnet wird, wie sie durch den genannten bekannten Merkmalsvektor definiert wird;
    • Erzeugen (120) einer Austauschklasse von Datenproben aus der genannten einzelnen Klasse von Datenproben, basierend auf dem Umgrenzungsradius unter Erzeugung einer gleichförmigen Verteilung von statistischen Austausch- Vektorelementen, welche einen Gruppenraum für die genannte Austauschklasse von Datenproben einnehmen, wobei die Vektorelemente aus einer gleichförmigen Verteilung ausgewählt sind, sowie unter Löschung aller Vektorelemente innerhalb einer Hypersphäre des genannten Umgrenzungsradius eines Merkmalsvektors aus der genannten einzelnen Klasse zur Erzeugung der genannten Austauschklasse; und
    • Trainieren (130) der genannten Klassifizierungseinrichtung auf der Basis der beiden Klassen, welche durch den Umgrenzungsradius definiert sind, wobei das Trainieren die Bestimmung einer Mehrzahl von Klassifizierungsgewichtungen umfaßt.
  3. Mustererkennungssystem (10), welches folgendes enthält:

    eine Klassifizierungseinrichtung, welche gemäß Anspruch 2 trainiert wird, um Rohdaten jeweils als Mitglied einer von zwei Klassen zuordnen, um die Erkennung eines Musters in den Daten zu erleichtern; und Mittel zur Anpassung des genannten abgeschätzten Umgrenzungsradius und der Klassifizierungsgewichtungen auf der Basis von Ergebnissen von der genannten Klassifizierungseinrichtung.
  4. Mustererkennungssystem nach Anspruch 3 in Anwendung auf die Erkennung von Mustern in einem Sonarsystem, wobei das Mustererkennungssystem weiter folgendes enthält:
    • Sendermittel (11) zur Erzeugung eines Sonarimpulses;
    • Ein Phased-Array-Strahlformungsnetzwerk (12) zum Empfang des Sonarimpulses zwecks Erzeugung eines Empfangssignales;
    • Signalverarbeitungsmittel (18) zum Detektieren des Empfangssignales und zum Extrahieren einer Mehrzahl von Datenproben daraus; und
    • Mittel (20) zum Extrahieren von Merkmalen der Datenproben.
  5. Verfahren zur Erkennung eines Musters in Datenproben mit folgenden Schritten:
    • Trainieren einer Klassifizierungseinrichtung gemäß Anspruch 1;
    • Klassifizieren (140) der Rohdaten jeweils als Mitglied einer von zwei Klassen zur Erleichterung der Erkennung eines Musters in den Daten; und
    • Anpassen des genannten abgeschätzten Umgrenzungsradius und der genannten Klassifizierungsgewichtungen auf der Basis laufender Ergebnisse aus dem Schritt der Klassifizierung der Rohdaten.
  6. Verfahren nach Anspruch 5 in Anwendung auf Sonardaten mit folgenden weiteren Schritten:
    • Erzeugung (11) eines Sonarimpulses;
    • Empfang (12) der Sonarimpulse zur Erzeugung eines Empfangsignales;
    • Detektieren (18) des Empfangsignales und Extrahieren einer Mehrzahl von Datenproben daraus; und
    • Extrahieren (20) von Merkmalen in den Datenproben.
  7. Fernbereichs- Sonarsystem mit verbesserter Mustererkennung zur Unterscheidung von Zielobjektdaten von Hintergrunddaten, wobei die genannten Hintergrunddaten eine Klasse nichtstationärer Störungsdaten bilden, mit einem Mustererkennungssystem nach Anspruch 3.
  8. Verfahren zur Unterscheidung von Daten gegenüber Hintergrunddaten in einem Sonarsystem, wobei die Hintergrunddaten eine Klasse nichtstationärer Störungsdaten bilden, mit der Ausführung der Verfahrensschritte nach Anspruch 6.
Anspruch[en]
  1. A method for training a classifier (100), including the steps of:
    • generating a plurality of feature vectors from a single class of exemplary non-stationary data samples:
    • estimating (110) a pruning radius based upon the distance between samples in the single class for said feature vectors in said single class of non-stationary data samples, said pruning radius being the radial distance from each feature vector beyond which an unknown feature vector presented for classification will be classified as outside said single class as defined by said known feature vector;
    • generating (120) a replacement class of data samples from said single class of data samples based on said pruning radius, including generating a uniform distribution of random replacement vector elements occupying a class space for said replacement class of data samples, wherein said vector elements are selected from a uniform distribution, and deleting all vector elements within a hypersphere of said pruning radius of a feature vector from said single class, to provide said replacement class; and
    • training (130) said classifier based on said two classes defined by said pruning radius, the training including determining a plurality of classifier weights.
  2. A pattern recognition system (10) comprising a classifier trained by the process of:
    • generating a plurality of feature vectors from a single class of exemplary non-stationary data samples:
    • estimating (110) a pruning radius based upon the distance between samples in the single class for said feature vectors in said single class of non-stationary data samples, said pruning radius being the radial distance from each feature vector beyond which an unknown feature vector presented for classification will be classified as outside said single class as defined by said known feature vector;
    • generating (120) a replacement class of data samples from said single class of data samples based on said pruning radius, including generating a uniform distribution of random replacement vector elements occupying a class space for said replacement class of data samples, wherein said vector elements are selected from a uniform distribution, and deleting all vector elements within a hypersphere of said pruning radius of a feature vector from said single class, to provide said replacement class; and
    • training (130) said classifier based on said two classes defined by said pruning radius, the training including determining a plurality of classifier weights.
  3. A pattern recognition system (10) comprising:
    • a classifier trained as claimed in claim 2 to classify raw data as a member of one of two classes to facilitate the recognition of a pattern in said data; and
    • means for adapting said estimated pruning radius and said classifier weights based on results from said classifier.
  4. The pattern recognition system, as claimed in claim 3, applied to the recognition of patterns in a sonar system, further comprising:
    • transmitter means (11) for providing a sonar pulse;
    • phased array beam forming network means (12) for receiving said sonar pulse to provide a received signal;
    • signal processing means (18) for detecting said received signal and extracting a plurality of data samples therefrom; and
    • means (20) for extracting features in said data samples.
  5. A method of recognising a pattern in data samples; comprising the steps of: training a classifying means as claimed in claim 1;
    • classifying (140) raw data as a member of one of two classes to facilitate the recognition of a pattern in said data; and
    • adapting said estimated pruning radius and said classifier weights based on current results from said step of classifying raw data.
  6. The method of claim 5 applied to sonar data, further comprising the steps of:
    • providing (11) a sonar pulse;
    • receiving (12) said sonar pulse to provide a received signal;
    • detecting (18) said received signal and extracting a plurality of data samples therefrom; and
    • extracting (20) features in said data samples.
  7. A long range sonar system having an improved pattern recognition capability to discriminate target data from background data, wherein said background data is a class of non-stationary clutter data, comprising: the pattern recognition system as claimed in claim 3.
  8. A method for discriminating data from background data in a sonar system, wherein said background data is a class of non-stationary clutter data, comprising: carrying out the steps as claimed in claim 6.
Anspruch[fr]
  1. Procédé pour l'apprentissage d'un classificateur (100), comprenant les étapes suivantes :
    • on génère une pluralité de vecteurs de caractéristiques à partir d'une classe unique d'échantillons de données non stationnaires pris à titre d'exemples;
    • on estime (110) un rayon d'élagage basé sur la distance entre des échantillons dans la classe unique pour les vecteurs de caractéristiques dans la classe unique d'échantillons de données non stationnaires, ce rayon d'élagage étant la distance radiale à partir de chaque vecteur de caractéristiques au-delà de laquelle un vecteur de caractéristiques inconnu présenté pour la classification sera classé comme étant à l'extérieur de la classe unique définie par le vecteur de caractéristiques connu;
    • on génère (120) une classe d'échantillons de données de remplacement à partir de la classe unique d'échantillons de données sur la base dudit rayon d'élagage, ceci comprenant la génération d'une distribution uniforme d'éléments aléatoires de vecteur de remplacement occupant un espace de classe pour la classe d'échantillons de données de remplacement, ces éléments de vecteur étant sélectionnés parmi une distribution uniforme, et la suppression de tous les éléments de vecteur à l'intérieur d'une hypersphère du rayon d'élagage d'un vecteur de caractéristiques provenant de la classe unique, pour fournir la classe de remplacement; et
    • on effectue l'apprentissage (130) du classificateur sur la base des deux classes définies par le rayon d'élagage, l'apprentissage comprenant la détermination d'une pluralité de poids de classificateur.
  2. Système de reconnaissance de formes (10) comprenant un classificateur dont l'apprentissage est effectué par le processus suivant :
    • on génère une pluralité de vecteurs de caractéristiques à partir d'une classe unique d'échantillons de données non stationnaires pris à titre d'exemples;
    • on estime (110) un rayon d'élagage basé sur la distance entre des échantillons dans la classe unique pour les vecteurs de caractéristiques dans la classe unique d'échantillons de données non stationnaires,' ce rayon d'élagage étant la distance radiale à partir de chaque vecteur de caractéristiques au-delà de laquelle un vecteur de caractéristiques inconnu présenté pour la classification sera classé comme étant à l'extérieur de la classe unique définie par le vecteur de caractéristiques connu;
    • on génère (120) une classe d'échantillons de données de remplacement à partir de la classe unique d'échantillons de données sur la base du rayon d'élagage, ceci comprenant la génération d'une distribution uniforme d'éléments de vecteur de remplacement aléatoires occupant un espace de classe pour la classe d'échantillons de données de remplacement, ces éléments de vecteur étant sélectionnés parmi une distribution uniforme, et la suppression de tous les éléments de vecteur à l'intérieur d'une hypersphère du rayon d'élagage d'un vecteur de caractéristiques provenant de la classe unique, pour fournir la classe de remplacement; et
    • on effectue l'apprentissage (130) du classificateur sur la base des deux classes définies par le rayon d'élagage, l'apprentissage comprenant la détermination d'une pluralité de poids de classificateur.
  3. Système de reconnaissance de formes (10) comprenant :
    • un classificateur dont l'apprentissage est effectué de la manière revendiquée dans la revendication 2, pour classer des données brutes comme appartenant à l'une de deux classes, pour faciliter la reconnaissance d'une forme dans ces données; et
    • des moyens pour adapter le rayon d'élagage estimé et les poids de classificateur sur la base de résultats du classificateur.
  4. Système de reconnaissance de formes selon la revendication 3, appliqué à la reconnaissance de formes dans un système de sonar, comprenant en outre :
    • des moyens émetteurs (11) pour fournir une impulsion de sonar;
    • une structure de réseau de formation de faisceau (12) utilisant un groupement à déphasage, pour recevoir l'impulsion de sonar afin de fournir un signal reçu;
    • des moyens de traitement de signal (18) pour détecter le signal reçu et en extraire une pluralité d'échantillons de données; et
    • des moyens (20) pour extraire des caractéristiques dans les échantillons de données.
  5. Procédé de reconnaissance d'une forme dans des échantillons de données, comprenant les étapes suivantes : on effectue l'apprentissage de moyens de classification de la manière revendiquée dans la revendication 1;
    • on classe (140) des données brutes comme appartenant à l'une de deux classes, pour faciliter la reconnaissance d'une forme dans ces données; et
    • on adapte le rayon d'élagage estimé et les poids de classificateur sur la base de résultats présents provenant de l'étape de classification de données brutes.
  6. Procédé selon la revendication 5, appliqué à des données de sonar, comprenant en outre les étapes suivantes :
    • on fournit (11) une impulsion de sonar;
    • on reçoit (12) l'impulsion de sonar pour fournir un signal reçu;
    • on détecte (18) le signal reçu et on en extrait une pluralité d'échantillons de données; et
    • on extrait (20) des caractéristiques dans les échantillons de données.
  7. Système de sonar à longue portée ayant une possibilité de reconnaissance de formes améliorée pour la discrimination entre des données de cible et des données de fond, dans lequel les données de fond sont une classe de données de fouillis non stationnaires, comprenant : le système de reconnaissance de formes revendiqué dans la revendication 3.
  8. Procédé pour discriminer entre des données et des données de fond dans un système de sonar, dans lequel les données de fond sont une classe de données de fouillis non stationnaires, comprenant : l'accomplissement des étapes revendiquées dans la revendication 6.






IPC
A Täglicher Lebensbedarf
B Arbeitsverfahren; Transportieren
C Chemie; Hüttenwesen
D Textilien; Papier
E Bauwesen; Erdbohren; Bergbau
F Maschinenbau; Beleuchtung; Heizung; Waffen; Sprengen
G Physik
H Elektrotechnik

Anmelder
Datum

Patentrecherche

Patent Zeichnungen (PDF)

Copyright © 2008 Patent-De Alle Rechte vorbehalten. eMail: info@patent-de.com