DE102009013390A1 - Method for automatic classification of physical object i.e. postal item, involves parameterizing classifier for classification of physical objects using parameter values that are computed by global optimization - Google Patents

Abstract

The method involves utilizing a preset weighting function as a local to-be-optimized target function. A classifier (2) is parameterized on a trial basis with different values for an influencing parameter. A global optimum for parameters (P1-Pn) is computed. The weighting function is utilized as a global to-be-optimized target function. The influencing parameter is varied under utilization of computed local optimums. The classifier is parameterized for classification of physical objects i.e. postal items, using parameter values that are computed by global optimization. Independent claims are also included for the following: (1) a data storage medium comprising a computer program for performing a method for automatic classification of a physical object (2) a device for automatic classification of a physical object.

Description

The The invention relates to a method and apparatus for classifying a physical object, in particular for reading a string on a mailing, by means of a classifier, first is parameterized and then a lot of sensor values evaluates from the object.

The Performance of such a classifier often depends critically which values have parameters of the classifier. These Parameter values remain unchanged during classifying.

In US 7,031,512 B2 A method for automatically recognizing characters is described. The method is z. B. used to automatically read an address on a mailpiece.

In a database ("directory retrieval system") are valid classification results stored, z. B. actually existing addresses. A digital image the mailpiece is evaluated. A classifier deciphers the Characters in this image or returns the result, the characters unable to classify. This classifier leads successively several steps through, z. "Image enhancement", "address block locator "," segmentation "," feature extraction "," character selection "and" confidence evaluation ". Every step depends on at least a parameter of the classifier. The classifier leads at least one classification run (one iteration) through while all parameters have constant values, that is not varied become. The classification result, which is the classification run supplies, on the one hand, with the valid classification results compared in the database. On the other hand, a confidence measure ("confidence level ") compared with a predetermined barrier. If the classification result is too uncertain or not with one valid result agrees, so will at least a parameter value changes, and a new classification run (a new iteration) is performed. These iterations are repeated until the classifier becomes a valid result delivers or a termination criterion is met, z. Legs predetermined maximum time has elapsed.

In A. Berlanga et al .: "Optimizing Statistical Character Recognition Using Evolutionary Strategies to Recognize Aircraft Tail Numbers", EURASIP Journal on Applied Signal Processing 2004: 8, pp. 1125-1134 , a method is described for decoding a tag on a tail of an aircraft. Here, too, a database with valid results is given.

Of the Invention is based on the object, a method and an apparatus by which an object is automatically classified is measured, whereby the object a lot of sensor values is measured and a classifier by applying an algorithm these sensor values evaluates and this algorithm of the values of various parameters depends on the parameters set manually without will need. The method and the device should also then be applicable when setting hundreds or even thousands of parameters are.

The The object is achieved by a method having the features of the claim 1 and a device with the features of claim 26 solved. Advantageous embodiments are in the subclaims specified.

According to the solution, a method and a device for classifying a physical object are provided. Classifying involves the following steps:
A sensor makes measurements on the physical object. As a result, a lot of sensor values are generated.

The The amount of sensor values generated by the measurement becomes more computer-available Form saved.

The physical object is to be classified. The default is one Set of classes as the possible classification results. Any classification result is a class or a vector of classes to which the object belongs.

One Classifier classifies the object. The classifier presents Determines to which of the possible classes the object belongs. A classification result provided by the classifier is for example, a class or vector of classes. Possible is also that the classifier yields the result that the Object belongs to several of the given classes. Possible is also that the classifier determines that the object does not belong to any one belongs to the given classes. It is also possible that the classifier returns the result, the object does not classify to be able to.

Of the Classifier acts as a pattern recognition device and uses the stored sensor value set for classification. For this the classifier applies an algorithm, to detect at least one pattern in the sensor value set.

• – entweder das Ergebnis der Auswertung der Sensorwerte-Menge
• – oder das Ergebnis, die Sensorwerte-Menge (Abb) nicht klassifizieren zu können.

The classifier returns one of the following two results as a classification of the object:

• Either the result of the evaluation of the sensor value quantity
• - or the result of not being able to classify the sensor value set (Fig.).

Of the depends on the algorithm used for classifying of several parameters. The values of these parameters remain during of classifying unchanged.

In front The classifier is automatically parameterised for classifying. The parameterization includes the step that each parameter of the Algorithm is assigned a value that previously determined automatically becomes.

• – Eine Objekt-Stichprobe mit N physikalischen Stichproben-Objekten wird vorgegeben.
• – Ein Sensor erzeugt eine Stichprobe mit N Sensorwerte-Mengen. Hierfür nimmt der Sensor Messungen an den N Stichproben-Objekten vor. Die N Sensorwerte-Mengen werden in rechnerverfügbarer Form abgespeichert.
• – Für jede Sensorwerte-Menge der Stichprobe wird jeweils ein Soll-Klassifikationsergebnis ermittelt und in rechnerverfügbarer Form abgespeichert.
• – Eine Bewertungsfunktion (Bew) wird vorgegeben. Diese Bewertungsfunktion (Bew) hängt von der Quote der falsch klassifizierten Sensorwerte-Mengen und der Quote der als nicht klassifizierbar eingestuften Sensorwerte-Mengen ab.

To automatically determine the values, the following steps are performed:

• - An object sample with N physical sample objects is specified.
• - A sensor generates a sample with N sensor value quantities. For this purpose, the sensor makes measurements on the N sample objects. The N sensor value quantities are stored in computer-available form.
• For each sensor value set of the random sample, a desired classification result is determined in each case and stored in computer-available form.
• - An evaluation function (Bew) is specified. This weighting function (Bew) depends on the proportion of misclassified sensor value quantities and the rate of sensor value sets classified as non-classifiable.

Around parameterizing the classifier will continue to be a first Step, a second step and a third step performed.

• – Der Klassifikator wird wiederholt probeweise mit unterschiedlichen Parameterwerten parametriert.
• – Der parametrierte Klassifikator wird auf jeweils mindestens eine Sensorwerte-Menge der Stichprobe angewendet.
• – Für jede dieser Sensorwerte-Mengen wird das Ist-Klassifikationsergebnis, das der Klassifikator liefert, mit dem abgespeicherten Soll-Klassifikationsergebnis für diese Sensorwerte-Menge verglichen.

In each of these three steps, the following steps are performed:

• - The classifier is repeatedly parameterized on a trial basis with different parameter values.
• - The parameterized classifier is applied to at least one sensor value set of the random sample.
• For each of these sensor value sets, the actual classification result which the classifier delivers is compared with the stored desired classification result for this sensor value set.

• – dem Ist-Klassifikationsergebnis, das der Klassifikator angewendet auf diese Sensorwerte-Menge liefert, und
• – dem abgespeicherten Soll-Klassifikationsergebnis für diese Sensorwerte-Menge

für verschiedene Werte des Parameters unterschiedlich ausfällt.In the first step, at least one influencing parameter of the algorithm, which influences at least one sensor value set of the random sample, is determined. In this case, a parameter then influences a sensor value quantity of the random sample if the result of the comparison between

• The actual classification result that the classifier applies to this sensor value set, and
• The stored nominal classification result for this sensor value set

varies for different values of the parameter.

• – Die vorgegebene Bewertungsfunktion wird als die lokal zu optimierende Zielfunktion verwendet.
• – der Klassifikator wird probeweise mit verschiedenen Werten für den beeinflussenden Parameter parametriert. Hierbei bleiben die Werte aller anderen Parameter unverändert.

In the second step, a local optimum is calculated for each influencing parameter determined in the first step. In order to calculate the local optimum of an influencing parameter in the second step, the following steps are carried out:

• - The default evaluation function is used as the target function to be locally optimized.
• - The classifier is parameterized on a trial basis with different values for the influencing parameter. The values of all other parameters remain unchanged.

• – Die vorgegebene Bewertungsfunktion wird als die global zu optimierende Zielfunktion verwendet.
• – Jeder beeinflussende Parameter wird unter Verwendung des im zweiten Schritt berechneten lokalen Optimums variiert.
• – Der Klassifikator wird für das Klassifizieren des Objekts mit denjenigen Parameterwerten parametriert, die bei der globalen Optimierung im dritten Schritt berechnet wurden.

In the third step, a global optimum for the parameters is calculated. To calculate the global optimum for the parameters, the following steps are performed in the third step:

• The predefined evaluation function is used as the target function to be optimized globally.
• Each influencing parameter is varied using the local optimum calculated in the second step.
• - The classifier is parameterized for classifying the object with those parameter values that were calculated in the third step in the global optimization.

The Solution-based method determines automatically those parameters that determine the classification performance of the classifier the strongest influence. This happens in particular because by that in the first step or the influencing parameters be determined. This approach is much more targeted and faster as a procedure in which all parameters are systematically varied and the effect of each variation is checked. This systematic approach is for a large number Parameters very time consuming.

The solution according to the method is also more targeted than the z. B. off US 7,031,519 B2 known method, which consists of successively varying at least one parameter until either a result classified as correct is found or an abort condition is fulfilled.

The Method according to the invention comprises a Parameterization phase and a classification phase. The parameterization phase consists of three steps and provides a parameterized classifier. This classifier is used in the classification phase. Only for the parameterization phase is a sample with nominal classification results needed. The steps of the parameterization phase need only once to be performed, and the parameterized Classifier can be used as often as you like. In the Classification phase need thanks to the invention no iteration runs to be performed with different parameter values, which saves time.

One Advantage of the method is that there are several process steps run in parallel. This saves computing time. Become a solution namely determined in the first step influencing parameters. The Check if a parameter is an influencing parameter is independent of and therefore parallel to the corresponding test for another parameter carry out. No exam for one Parameter requires the result of the check for another parameter. Also the tests, which sensor value quantities be influenced by a parameter and which not, let to execute in parallel.

in the second step, each influencing parameter is locally optimized. These calculations can also be performed in parallel. Because the local optimization of a parameter does not require that previously another parameter of the algorithm was optimized.

The Number of computation steps used in the first step of the solution To carry out the process (influenced sensor value quantities), depends only linearly on the number of parameters, not about square or exponential. The same applies to the Number of calculation steps in the second step (local optimizations).

in the third step, the sensor value set is evaluated to a global one To find the optimum. Also with this optimization can be several Perform steps in parallel.

Out These reasons can be the solution Use procedures even when setting many parameters are, for example, several hundred or even a thousand parameters. In this case it would not be possible to set the parameters set manually. The optimization method according to the state of Technology usually requires a much too high computation time and are therefore not practical. In many applications, too the available computing capacity is not enough. The Solution-based method avoids these bottlenecks.

The solution according to the method leaves to apply to an algorithm with arbitrary parameters. These Parameters can be any number of different value ranges, Have units of measurement and physical meanings.

Preferably Not every parameter is varied in the second and third steps. Rather, in the first step it is determined which parameters the Classification performance of the classifier can influence at all. In the first step, the set of parameters becomes systematic the influencing parameters selected. The remaining Parameters remain constant. This reduces the computational effort for the subsequent steps clearly.

One Parameter is then selected as the influencing parameter if the classifier at different values of this parameter for at least one selected sensor value set gives different results. Preferably, the predetermined Initial value of the parameter of one of these different values. The unaffected parameters retain their respective predetermined ones Initial values.

Prefers For this purpose, first those sensor value quantities are selected that does not classify correctly by the classifier if the classifier is parameterized with given initial values is. Subsequently, the influencing parameters selected. This is the sample parameterized classifier applied exclusively to those sensor value quantities, that of the classifier parameterized with the initial values not can classify correctly.

These Design reduces the number of those parameters at the parameterization can be varied dramatically. Spared at the same time the design the need to manually vary the ones Specify parameters. Rather, it will be automatic and systematic selected which parameters are varied and which Not.

Preferably will take the sample with the sensor value sets in sub-samples divided. The calculation steps of the first step are for each sub-sample performed. The partial samples can be temporally parallel or at least temporally overlapping process, so that several computers can be used and the Parallelization saves time.

The solution according to the method leaves z. B. to use to digitize image sequences and thereby to reconstruct. An analog film, z. An old feature film, is optically scanned. The scan provides computer-accessible Image sequences that are still with mechanical or chemical damage the film may be afflicted. A reconstruction method applies an embodiment of the method according to the solution and provides error-free and reconstructed image sequences. One Desired classification result for a sample obtained by sampling Picture sequence is a correctly reconstructed picture sequence.

In another application of the solution Procedure will automatically for a picture sequence at least generates a descriptor describing the image sequence ("video annotation"). The image sequence was z. B. generated by a section of the real World was filmed, z. For example, a parliamentary debate or a sporting event. The image sequence is z. B. in analog form in the archive of a TV station and is digitized by sampling, which is at least provides a sensor value quantity. The descriptor is z. B. one of the terms "Bundestag debate", "football match" or "natural disaster". As the target classification result for an image sequence The correct descriptors are determined and stored.

In a modification, the image sequence is obtained by a technical System is monitored by means of at least one camera. The technical system is z. B. a production plant, a building or a busy street, a Airport or other public place. An embodiment of the inventive method is on applied the image sequence and provides a statement about a current or future state of the technical Systems, z. As a threat to people or an impending Disruption of the plant or an accident or a traffic jam on the Street.

In Another application is the solution according to Method used to draw strings on objects to read. Each string is z. B. a delivery address on a Mailing or on a freight, the name of the owner of one Baggage or the official license plate a road, rail, air or water vehicle. The indicator should z. B. can be determined for a fee for the use of a road to raise or to investigate an administrative offense.

For example The procedure is used to both the destination address and recognize the address of the sender and also to check if the mailing with a sufficient Transport fee is franked (valid and sufficient stamps). The classification result is a vector with three components, namely the destination address, the sender address and the recognized fare. Possible is z. For example, although the destination address and the transportation fee be recognized correctly, but not the sender address.

From an object which is provided with a string becomes generates a computer-accessible image. This image shows the string. The target classification result is an encoding saved and used the correct read string.

The solution according to the method leaves also use to classify an object. From the object At least one computer-accessible image was generated. The object is determined by its shape, colors and / or its Dimensions classified in the image. The image is z. B. by An imaging method of medicine measures and displays body parts of a human being or another living thing.

The object to be classified can also be an object, eg. A vehicle, a component, a workpiece or a returnable bottle. The object may also include a stamp, a bill, an identity card, a credit card or other valuable document that checks for authenticity shall be. A computer-accessible library is specified with images of real value documents. To classify a value document to be checked optically scanned. The classification result is either a value document from the Library or the result that may be a Counterfeiting is present, delivered.

The Method is also able to evaluate geo-information. As sensor value quantity becomes a computer-accessible image of a section of the Earth's surface used in lighting in the visible Range or generated by infrared. This section of the earth's surface should be classified. For example, underground minerals are found, the state of green areas classified or environmental pollution, eg. B. oil films on waters, be recognized.

The Sensor values can also be supplied by an acoustic sensor be that measures noise. The sounds can come from a technical system and are classified with the goal to recognize the current state of the technical system. Certain Noises can be on or wear indicate a damage. The technical system is z. B. a conveyor system for cargo or mineral resources, an engine, a Transmission or a chassis of a vehicle.

The Method can also be used for speech recognition apply. A sequence with at least one spoken word becomes recorded by an acoustic sensor. This sensor value sequence becomes with the aim of automatically adding the spoken words detect. For example, a worker speaks control commands or observations in a head-worn microphone to the commands or observations to enter and to have both hands free.

In Another application is to regulate a technical system. For example, should a vehicle or an airplane on a given Trajectory are kept, even if disturbances trying to get it off the track. Speed is measured Acceleration, orientation and other physical quantities. The target classification result is a correct control intervention in response to measured sensor values and / or specifications of reference variables used.

in the The invention is based on an embodiment described. Showing:

1 a sorting plant with a classifier;

2 a detailed view of the classifier of 1 ;

3 a parameterization unit containing the database of the classifier of 1 provides with parameter values.

4 an embodiment of that other computer that performs the parameterization, with a central computer and two satellite computers;

5 an illustration of the preparatory steps in which the N images and the N target reading results are determined and stored;

6 an illustration of the preparatory step, in which the actual reading result of the parameterized with the initial values classifier is determined;

7 an illustration of the preparatory step, in which it is determined which images are evaluated by means of which parameter values;

8th an illustration of the preparatory step in which the N1 not completely correctly read images are determined;

9 an illustration of how the influencing parameters are determined;

10 an illustration of how the affected images are detected;

11 an illustration of how the influencing parameters are optimized locally;

12 an illustration of how to optimize globally.

in the Embodiment is the solution according Methods used to detect strings on mail. The mail items are sorted by a sorting system, the after sorting the transport of each mail item to the respective one Target triggers. Each mail item is each with at least one String provided. The string is z. B. the delivery address, to which the item of mail is to be transported, a return address or an advance or transportation request ( "Endorsement").

• – einer Zuführeinrichtung ZE,
• – mehreren Ausgabefächern Af-A, Af-B, ..., Af-E,
• – einer Kamera 3 und
• – einem Klassifikator 2.

1 shows a sorting system 1 with the following components:

• A feeder ZE,
• - several output trays Af-A, Af-B, ..., Af-E,
• - a camera 3 and
• - a classifier 2 ,

A stack of mail items is sent via the feeder ZE of the sorting system 1 fed. The feeder ZE separates the mail items, which then as stream of isolated mail the sorting system 1 run through. The camera 3 the sorting plant 1 generates a respective computer-accessible image of each mail item in such a way that the image contains an image of the string on the mailpiece. This digital image is broken down into its components. Example becomes in 1 a mail item Ps with a delivery address XYZ shown by the camera 3 An image is generated.

The classifier 2 evaluates the image Abb and decodes the delivery address XYZ. The sorting system 1 uses this reading result LE and feeds the mail item Ps into one of the output bins Af-A, Af-B, ..., Af-E.

2 shows an example of a detailed view of the classifier of 1 , In this detail view, the classifier includes 2 five individual single classifiers 2.1 . 2.2 , ..., 2.5 which are connected in series. Each individual classifier uses an algorithm Al.1, Al.2, ..., Al.5 and delivers a result E.1, E.2, ..., E.5. The single classifier 2.2 uses the result E.1 from the previous classifier 2.1 , the single classifier 2.3 the result 2.2 and so on.

Each of these five single classifiers 2.1 . 2.2 , ..., 2.5 has parameters that are determined in other ways than by the method according to the solution. In addition, each individual classifier is influenced by the respective value of at least one of the parameters P (1), P (2), ..., P (n).

• – den Namen des Adressaten,
• – eine Hausnummer,
• – einen Straßennamen,
• – einen Ort,
• – ein Kürzel für den US-Bundesstaat und
• – einen ZIP-Code, der z. B. aus 5 oder 9 oder 11 Ziffern besteht.

In the exemplary embodiment, the method is used for mailings to addressees in the USA. Therefore, the image contains

• - the name of the addressee,
• - a house number,
• - a street name,
• - a place,
• - an abbreviation for the US state and
• - a ZIP code, the z. B. consists of 5 or 9 or 11 digits.

A character reading unit (OCR unit) of a classifier 2 the sorting plant 1 then returns a read string, e.g. B. the delivery address. Each part of the image of the delivery address is decoded for this purpose.

Of the Deciphered address component is compared with an address database. This address database contains valid delivery addresses in the USA. The address database is tree-shaped. Each city is the abbreviation of the respective state and the valid ZIP codes for delivery addresses assigned in this city. Each ZIP code is valid Street names for delivery addresses associated with this ZIP code.

The following table shows an example. A mailing is provided with the following delivery address: Address part string Name of the addressee John Doe House number 217 street name Eigth Ave. place Burnham State abbreviation PA ZIP code 17009

Around automatically reading the address automatically becomes a computer-available Generates an image of the mail item showing the address. In this Image is searched for the area containing the address ("Region of interest"). The area becomes integral decomposed to find the address components.

In In one embodiment, the place name first becomes searched and deciphered. The deciphering delivers in the present Example not a clear result, but several possible ones Reading results with one match measure each (confidence measure). This confidence measure is a measure of how good the reading result for the address part with the respective Address part of a valid address in the address database matches. Each agreement measure is in the embodiment a number between 0 (lowest result) and 255 (highest Result).

By Rather, comparison with the address database is determined that the place either "Burnham" (metric 223) or "Burningham" (Conformity measure 210) reads. All other place names stored in the address database fit significantly worse to the reading result.

to Control is reading and comparing the state abbreviation. "Burnham" includes "PA" (score 200), to "Burningham" "GA" (Conformity measure 210).

Subsequently becomes the reading result for the ZIP code with all ZIP codes stored in the address database for delivery addresses filed in "Burnham" or "Burningham" are. The only stored ZIP code that matches "Burnham", is "17009" (metric 130), the only ZIP code to match "Burningham" is "77008" (metric 120).

The same is done for the road. Then the house number is deciphered.

A Modification of this embodiment is performed if more runtime is available. First the ZIP code is searched for and deciphered. All deposited valid Addresses whose ZIP code suffices to match the reading result for the ZIP code are examined as described above.

The following table illustrates which two alternative delivery addresses were read as possible read results: Address part 1st alternative Single Review 2nd alternative Single Review place Burnham 223 Burningham 210 contraction PA 200 GA 210 ZIP code 17009 130 77008 120 street name Eighth Ave. 240 Eigth Ave. 255 House number 217 255 211 200

In the second and fourth columns are two valid alternatives (two delivery addresses listed in the address database). In the third and fifth column is as single rating each the measure of conformity (confidence measure) for the local match between the reading result for the address part and the valid string specified in the address database.

For every valid alternative (stored in the address database) An overall score is determined by the individual match sizes weighted to a total security measure (total confidence measure) be summarized. The place receives a higher one Weighting as the house number because the right place for the delivery is more important.

If the overall rating of each valid alternative below a given security threshold or at all no valid alternative is found, so deliver the Character reading unit the result that they are the string at all unable to decipher. In this case, the image of the mailpiece transmitted to a video coding station and on a video display device shown. An editor reads the mapped string, z. As the delivery address, and enter them manually, z. B. by Typing or voice input. This is of course more time consuming and more expensive than an automatic reading.

If Several valid alternatives are found whose ratings above the given threshold, the distance between compared these ratings with a distance threshold. If the distance is greater than the distance threshold, Thus, the character reading unit provides a unique result.

If the classifier 2 In this way, it can not provide a clear result, so it looks for address components that are included in all valid alternatives. These address components are provided by the classifier 2 as a result. In general, these common address components will be correct. However, the classifier is capable 2 not fully and correctly recognize the address.

The classifier 2 uses an algorithm Al to read the string. This algorithm Al for reading a string - and thus the classifier 2 N uses different parameters P (1),..., P (n), n> = 1. In the exemplary embodiment n can be about 300, ie up to 300 parameters are to be set, ie about 300 parameters are each a value assign. Also possible is a significantly higher number of parameters. The solution according to the method provides a way to set these n parameters and thus the classifier 2 to parameterize.

The classifier 2 preferably comprises a software program implementing the Al algorithm and a computer on which this software program is installed which executes the program. In 1 becomes a calculator 6 shown that to the classifier 2 belongs. This calculator 6 has read access to a data store 5 in which the algorithm Al with the n parameters P (1), ..., P (n) is stored in executable form. Preferably, the data memory is located 5 in the calculator 6 ,

In the embodiment, the classifier reads 2 As stated above, not only the string, but additionally calculates a total confidence measure (total confidence score). This overall safety measure is a measure of what the safety of the classifier 2 Delivered result matches the string on the mailpiece, so the reading result is correct. The classifier 2 returns the read string if the calculated safety measure is greater than or equal to the safety threshold and if all other possible reading results have a significantly lower overall safety margin. "Significantly less" means: the distance is greater than a predetermined distance barrier, whereby this distance barrier is a parameter of the algorithm Al. Otherwise, the classifier delivers 2 the result is that the string can not be read with sufficient certainty. A low value for the safety threshold and a low value for the distance barrier lead to many incorrect read results, high values to many unreadable strings.

This safety threshold is in the embodiment not a fixed value, but a variable parameter, for by the parameterization of the classifier 2 a value is determined. During parameterization, the security threshold is repeatedly assigned a value on a trial basis and used for classification. The distance threshold described above is also a variable parameter. It is also possible that the individual match dimensions of the address components are combined in different ways to the overall security measure and the type of summarization, ie the aggregation rule, is a variable parameter.

It is also possible to use different safety thresholds. Every security threshold is a parameter. To deliver a reading result, select the classifier 2 First by a preselection reading results and determines how many reading results are preselected. The classifier 2 also selects one of the predetermined security thresholds, depending on the number of preselected read results. With only one preselected read result, the classifier chooses 2 one with a low value, because the only possible reading is probably correct. For many preselected read results, the classifier chooses 2 a threshold with a large value, because the risk of a false reading result is great. This embodiment thus leads to several parameters for the safety threshold.

Another parameter is used to convert a color image or gray image into a black-and-white image. Each image of a mailpiece is present as a pixel image. The scanning of the mailpiece provides for each pixel each a color value or a gray value, z. For example, a number between 0 and 255 (inclusive). A parameter of the algorithm Al and thus of the classifier 2 is the threshold for when a color value or gray value is converted to black. Not necessarily takes every parameter of the classifier 2 only integer numbers as values. The method can likewise be used for parameters whose values are decimal numbers (eg "float").

• – das Auffinden desjenigen Bereichs im Abbild der Postsendung, in dem sich die Zustelladresse befindet („region of interest”),
• – die Unterteilung einer Zeichenfolge in die einzelnen Worte („Segmentierung”),
• – die Unterteilung eines Worts in Buchstaben,
• – die Entzifferung von Handschriften.

Further parameters of the classifier 2 Control the following substeps when reading a string:

• Finding the region in the image of the mail item in which the delivery address is located ("region of interest"),
• - the subdivision of a string into the individual words ("segmentation"),
• - the subdivision of a word into letters,
• - the deciphering of manuscripts.

• – der Rate von richtig gelesenen Zeichenfolgen auf Postsendungen,
• – der Rate von Zeichenfolgen, die nicht automatisch gelesen werden konnten,
• – den Kosten, die durch das fehlerhafte Lesen einer Zeichenfolge entstehen, und
• – den Kosten, die dadurch entstehen, dass eine Zeichenfolge nicht automatisch gelesen werden konnte, sondern das Abbild an die Videocodierstation übermittelt werden musste.

A rating function to be optimized is specified. Values are assigned to the parameters such that the assigned values maximize or minimize this default rating function. The valuation function depends on the following key figures:

• The rate of correctly read strings on mailpieces,
• The rate of strings that could not be read automatically
• The costs of incorrectly reading a string, and
• - The costs that result from the fact that a string could not be read automatically, but the image had to be transmitted to the video coding station.

For example, the evaluation function has the form α · Ri - β · NL, where Ri is the percentage of correctly read strings, NL is the percentage of strings classified as unreadable, and α> 0 and β> 0 are two default weight factors.

• 1) Der Ort in der Adresse wurde falsch gelesen.
• 2) Der Ort wurde richtig gelesen, der ZIP-Code (oder die Postleitzahl) wurde aber falsch gelesen oder konnte nicht automatisch entziffert werden.
• 3) Der Ort und der ZIP-Code bzw. die Postleitzahl wurden richtig gelesen, aber die Straße wurde falsch gelesen oder konnte nicht automatisch entziffert werden.
• 4) Nur die Hausnummer oder die Postfachnummer wurden falsch gelesen, die übrigen Bestandteile der Adresse hingegen richtig.

If the string is an address, different types of read errors are distinguished because the different types cause differently serious errors in the delivery of the mail item:

• 1) The location in the address was read incorrectly.
• 2) The place was read correctly, but the zip code (or zip code) was misread or could not be deciphered automatically.
• 3) The location and ZIP code / postal code were read correctly, but the street was misread or could not be deciphered automatically.
• 4) Only the house number or PO box number were read incorrectly, the other components of the address, however, correct.

These types of errors are of varying severity because they cause different levels of cost and transport time. If the mail item is sent to a wrong place, it must be subsequently transported to the right place, which is expensive and costs a lot of time. Has the classifier 2 however, only the house number wrongly read, so the deliverer can correct this error during his tour.

wobei

• – s verschiedene Arten von Adress-Bestandteile unterschieden werden, z. B. Ort, ZIP-Code, Straße und Hausnummer,
• – r verschiedene Arten von Lesefehlern unterschieden werden (k = 1, ..., r),
• – NL der Prozentsatz der als nicht lesbar klassifizierten Zeichenfolgen ist,
• – β > 0 ein vorgegebener Gewichtsfaktor ist,
• – Ri(l) der Prozentsatz der richtig gelesenen Adress-Bestandteile ist (l = 1, ..., s),
• – LF(k) der Prozentsatz der Lesefehler der Art k an den insgesamt gelesenen Zeichenfolgen ist (k = 1, ..., r),
• – α(l) ein vorgegebener Gewichtsfaktor für den Adress-Bestandteil l (l = 1, ..., s) ist und
• – γ(k) ein vorgegebener Gewichtsfaktor für die Fehlerart k ist (k = 1, ..., r).

In a further education the evaluation function has the following form:

in which

• - s different types of address components are distinguished, for. Place, ZIP code, street and house number,
• - different types of read errors are distinguished (k = 1, ..., r),
• - NL is the percentage of strings classified as unreadable,
• Β> 0 is a given weight factor,
• - Ri (l) is the percentage of correctly read address components (l = 1, ..., s),
• - LF (k) is the percentage of read errors of type k on the total read strings (k = 1, ..., r),
• - α (l) is a predetermined weighting factor for the address component l (l = 1, ..., s) and
• - γ (k) is a given weight factor for the error k (k = 1, ..., r).

To the algorithm Al and thus the classifier 2 to parameterize, become a parameterization unit 4 used and a sample with N mail Ps (1), ..., Ps (N) predetermined. In each case, a string is printed on each of these N mailings, also a unique identifier for the mailing, z. B. coded in the form of a bar pattern ("bar code"). The character sequence is preferably the delivery address to be recognized, with which the item of mail is provided.

In one embodiment, N2 mailpieces (0 <N2 <N) of the sample are identified as such mail items as are from the classifier 2 definitely have to be read correctly. The classifier 2 is therefore to be parameterized such that it delivers an actual read result for each of these mailpieces that matches a desired read result.

The classifier 2 In a classification phase, decipher correctly characters on an image Abb, as in 1 will be shown. In a preceding parameterization phase the classifier becomes 2 parameterized.

3 illustrates how in the parameterization phase a parameterization unit 4 that the database 5 of the classifier 2 from 1 with optimal parameter values Opt (1), ..., Opt (n) for which n provides parameters P (1), ..., P (n). The parameterization unit 4 may be in a different location than the sorting facility 1 are located. It is possible that a parameterization unit 4 the n parameter values are determined and several classifiers that work in sorting systems in different sorting centers use these parameter values.

• – eine weitere Kamera 9,
• – einen weiteren Rechner 7, der die im Folgenden beschriebenen Verfahrensschritte ausführt, und
• – eine Stichproben-Datenbank 8 und
• – jeweils eine Parameterwerte-Zuweisungseinheit 13 pro Sortieranlage mit Klassifikator 2.

The parameterization unit 4 includes in the embodiment the following components:

• - another camera 9 .
• - another computer 7 which performs the method steps described below, and
• - a sample database 8th and
• Each one parameter value assignment unit 13 per sorting plant with classifier 2 ,

The parameter values assignment unit 13 assigns to each classifier those optimum parameter values Opt (1), ..., Opt (n), which are the parameterization unit 4 has determined.

In the classification phase, it is sufficient, the classifier 2 to realize with the help of a single computer. In the preceding parameterization phase the classifier becomes 2 preferably realized with the aid of a central computer and several satellite computers in order to save computing time. On each satellite machine, the algorithm is Al and a database 5 with the parameters installed. The steps of this parameterization phase are run through once, and the steps of the classification phase are performed for each object, e.g. B. each mail item, carried out again. The central computer and the satellite computers are connected to one another via a local computer network ("local area network"). The central computer calls the satellite computers. These deliver their results to the central computer, which processes these results.

4 schematically shows an arrangement with a central computer 7 and p = 2 satellite computers 6.1 and 6.2 , On the two satellite computers 6.1 and 6.2 is each a database 5.1 respectively. 5.2 with the algorithm Al and the n parameters P (1), ..., P (n) installed. This will create two classifiers 2.1 and 2.2 realized. These two classifiers 2.1 and 2.2 can be parameterized differently in the parameterization phase.

5 illustrates how a preparatory step S1 is performed. In this preparatory step S1, a computer-accessible image is generated and stored from each mail item of the random sample, which shows the character string on the mailpiece. Overall, the other camera produces 9 Thereby N images Abb (1), ..., Abb (N) with the N strings X [1], ..., X [N] on the N mailings Ps (1), ..., Ps (N) , These N images are stored in a data memory, in this case in the sample database 8th ,

In a further preparatory step S2, at least one desired reading result is determined and stored for each mail item, for example by first reading each character string automatically and then displaying the reading result on a screen, and an operator reads and inputs the character string. The classifier 2 should read the string one of the stored target reading results deliver.

It is possible that several desired reading results are stored. For example, a desired reading result for a handwritten delivery address is the completely correct read delivery address. Another target read result is the correct read ZIP code, with the result that the remainder of the delivery address can not be read automatically. Preferably, one of the desired reading results is characterized as the best possible desired reading result - usually the completely correct read string. In the example of 5 is a target reading result the completely correct read string X [j] on the mail item Ps (j) (j = 1, ..., N).

In the example of 3 is the mail item Ps (1) of the sample with the identifier K (1), the mail item Ps (2) with the identifier K (2), ..., the mail item Ps (N) with the identifier K (N). A target reading result for Ps (1) is X [1], which for Ps (2) is X [2], ..., which for Ps (N) is X [N].

• – das rechnerverfügbares Abbild Abb(j) der Postsendung Ps(j), das die Zeichenfolge X[j] zeigt,
• – die eindeutige Kennung K(j) der Postsendung Ps(j), die in codierter Form auf der Postsendung Ps(j) aufgedruckt ist, und
• – jeweils eine Codierung jedes Soll-Leseergebnisses für diese Postsendung.

In a table tab of the sample database 8th For each mail item Ps (j), one data record is created (j = 1, ..., N). Each record for a sample of the sample includes

• The computer-accessible image Abb (j) of the mail item Ps (j) showing the string X [j]
• - The unique identifier K (j) of the mail item Ps (j), which is printed in coded form on the mail item Ps (j), and
• - One coding of each desired reading result for this mail item.

The Identifier K (j) thus connects the mail item Ps (j) with the data record for the mail item Ps (j).

In the sample database 8th the N images Abb (1), ..., Abb (N) are stored, furthermore a table Tab with one data record per mailing. The record for the mail item Ps (l) contains the identifier K (l), the target reading result X [l] and a reference to the image Abb (l) (l = 1, ..., N).

each Image Abb (1), ..., Abb (N), the images stored to the N belongs, acts as a sensor value set of the sample.

For the method according to the solution, for each parameter P (i) of the classifier 2 in each case an initial value Ini (i) and a value range W (i) predetermined (i = 1, ..., n). Each value assigned to the parameter P (i) must be in this value range W (i) to be an allowable value. For a parameter P (i) (i = 1,..., N), for example, the value range from 0 to 255 (only integers) and the initial value 196 are specified. For example, both the initial value Ini (i) and the value range W (i) of a parameter P (i) are specified manually. It is also possible to use the result of a previous optimization as the initial value Ini (i).

In a further preparatory step it is determined which results of the classifier 2 then, if the classifier yields 2 is parameterized with the initial values Ini (1), ..., Ini (n) for the n parameters P (1), ..., P (n). The classifier 2 is applied to each stored image Fig. (l) of the sample (l = 1, ..., N). This application yields an actual read result I [l] for the image Fig. (L), in total N actual read results Is [1], ..., Is [N].

6 illustrates how the classifier parameterized with the initial values 2 evaluates the N images. As a result, on the one hand, the actual classification results delivered, on the other hand, a protocol Prot.

In this further preparatory step, therefore, the classifier 2 applied to all N images Abb (1), ..., Abb (N) of the sample. The classifier 2 applies the algorithm Al, which is realized as a software program. In one embodiment, a "tracing function" is used that some programming languages offer. The tracing function generates the operating log Prot. By evaluating this operating log Prot, it is determined for each image Abb (l) and each parameter P (i) whether the algorithm Al uses the value of the parameter P (i) to obtain the image l) or not, where l = 1, ..., N and i = 1, ..., n. For example, the tracing function automatically logs whether at least one program block in which P (i) occurs is the software Program that realizes the algorithm Al, while Fig (l) is evaluated. Whether such a program block is run through may depend on the properties of the image Fig. (L).

As a result, the quantity V (i) of images is determined for each parameter P (i) for their evaluation the algorithm Al uses the value of the parameter P (i), where the algorithm is Al and thus the classifier 2 are parameterized with the initial values. This set V (i) may be empty, ie the parameter P (i) is not used to evaluate even only one of the N images. Preferably, V (i) is generated and stored in the form of an index set, ie as a subset of the set {1, 2,..., N}.

7 illustrates the preparatory step in which is determined which images are evaluated using which parameter values.

In a subsequent preparatory step, the incorrect or at least incompletely read images of the random sample are determined. 8th illustrates this further preparatory step. In this preparatory step becomes the classifier 2 , which is parameterized with the initial values Ini (1), ..., Ini (n) for the n parameters P (1), ..., P (n). Those N1 images are selected for which the actual reading result If [1] either does not coincide with any of the stored desired reading results or at least not with the best possible set reading result X [1] - this is usually the completely correct read string. As a rule, only a small subset of all stored images is selected. For example, N is 100,000, and N1 is 2,000. This preparatory step provides N1 images [i (1)], Fig [i (2)], ..., Fig [i (N1)].

As just described, the classifier 2 in the parameterization phase preferably implemented by a central computer and p satellite computer. In this embodiment, the sample with the N images Abb [1], ..., Abb [N] is subdivided into p subsets. The k-th satellite calculates a subset of N (k) images (k = 1, ..., p), where N (1) + ... + N (p) = N. As a result, the k-th satellite computer in this preparatory step N1 (k) determines erroneous or at least incompletely read character strings. It is N1 (1) + ... + N1 (p) = N1. This embodiment is advantageous if the preparatory step requires a lot of computing time.

Possible But also, the N1 is read incorrectly or incompletely Images from a single computer, namely then, if the splitting and synchronization takes more time as saved by the parallelization.

In the three steps of the parameterization, at least one character string is repeatedly read on a trial basis. Each repeated read string is one of the N strings on one of the N mailings Ps (1), ..., Ps (N) of the sample. Here the classifier evaluates 2 This stored image decodes and decodes the string in the image - or provides the result that the string is not decipherable by machine. The classifier 2 is applied to the same image several times, each parameterized with different parameter values. This application returns an actual classification result in the form of an actual read result.

each Actual reading result is with the at least one desired reading result compared that in the record for this mailing is included.

• – Zeichenfolge vollständig richtig gelesen,
• – Zeichenfolge nur unvollständig gelesen, der gelesene Teil ist richtig,
• – Zeichenfolge wenigstens teilweise falsch gelesen,
• – Zeichenfolge konnte nicht automatisch gelesen werden („reject”).

In the exemplary embodiment, each comparison yields exactly one of the following results:

• - string read completely correctly,
• - string is only read incompletely, the part read is correct,
• - string at least partially misread,
• - String could not be read automatically ("reject").

As mentioned above, will be different types of reading errors if the string is an address. These different Types of read errors lead to further different ones Results that the comparison can have.

In the exemplary embodiment, first influencing parameters are selected in the first step. Only these influencing parameters are varied and optimized in the course of the parameterization. The remaining parameters retain their respective initial value Ini (i) as a value. 9 illustrates how the influencing parameters are determined.

As mentioned above, in one embodiment, a tracing function is used in the further preparatory step. For each parameter P (i), the set V (i) is determined, which consists of those images of the sample for the evaluation of which the classifier 2 the parameter P (i) used - more precisely: the initial value Ini (i) of P (i). If for a parameter P (i) this set V (i) is empty, then this parameter does not become is therefore automatically classified as a non-influencing parameter. In the further course of the parameterization phase and also in the classification phase, P (i) retains its initial value Ini (i) (i = 1,..., N).

Around the influencing parameters are selected in the first Step for each parameter the N1 selected Images Abb [i (1)], Abb [i (2)], ..., Abb [i (N1)] of the mail evaluated (m + 1) times to read the string. in the Read pass for the parameter P (i) (i = 1, ..., n) every other parameter P (j) (j = 1, ..., n; j # i) to the given Initial value Ini (j) is set. For parameter P (i), the Initial value w [i, 0] ≔ Ini (i) and m various others Values w [i, 1],..., W [i, m] are given from the value range W [i] of P (i). The m further values are z. B. selected by m / 2 Values greater and m / 2 values less than the initial value to be selected.

For each of these m values for the parameter P (i) becomes the classifier 2 on the N1 selected images Fig [i (1)], Fig [i (2)], ..., Fig [i (N1)] of the sample applied. These N1 selected images are those for which the classifier parameterized with the initial values Ini (1), ..., Ini (n) 2 provides an actual read result that deviates from the best possible target read result, that is to say incorrect or at least incomplete.

Will the classifier 2 parameterized with the value w [i, k] for the parameter P (i) (k = 0, 1,..., m), then it delivers the actual read results Is [i (1), i, k], If [i (2), i, k], ..., Ist [i (N1), i, k] for the N1 selected images Abb [i (1)], Abb [i (2)], .. ., Fig [i (N1)] (k = 0, 1, ..., m).

• – Das Ergebnis des Vergleichs zwischen dem jeweiligen Ist-Leseergebnis (der gelesenen Zeichenfolge), das der Klassifikator 2 liefert, und dem abgespeicherten bestmöglichen Soll-Leseergebnis X[i(l)] (der richtigen Zeichenfolge), ist für jeden der m + 1 untersuchten Werte des Parameters P(i) dasselbe.
• – Das Ergebnis des Vergleichs fällt für verschiedene Werte von P(i) unterschiedlich aus, die Ergebnisse des Klassifikators 2 hängen also bei diesem Abbild einer Postsendung davon ab, welchen Wert der Parameter P(i) hat.

For each selected image Abb [i (l)] (l = 1, ..., N1) two cases are distinguished:

• - The result of the comparison between the respective actual read result (the read string), the classifier 2 and the stored best-possible read-only result X [i (l)] (the correct string), is the same for each of the m + 1 examined values of the parameter P (i).
• The result of the comparison differs for different values of P (i), the results of the classifier 2 In the case of this image of a mail item, therefore, the value of the parameter P (i) depends on it.

In the illustration of 9 Δ [i (l)] denotes the result of the comparison between the actual reading result Ist [i (l)] (classifier 2 is parameterized with the initial values) and the best possible set reading X [i (l)]. Δ [i (l), i, k] denotes the result of comparison between the actual read result Is [i (1), i, k] (parameter P (i) has the value w [i, k]) and best possible set reading X [i (l)].

One Parameter P (i) is selected as an influencing parameter, if the result of the comparison for this parameter and selected for at least one image among the N1 Images Fig [i (1)], Fig [i (2)], ..., Fig [i (N1)] for different Parameter values for P (i) varies differently. Returns the comparison for each of the N1 images and for each value of the parameter examined the same Result, the parameter does not become an influencing parameter treats and keeps its initial value as a value.

First, the actual read result Is [i (j)] is determined, which is the classifier 2 then returns, if the classifier 2 is parameterized with the given initial values Ini (1), Ini (2), ..., Ini (n) and applied to an image. Subsequently, a further value w [i, k] is assigned to the parameter P (i) one after the other, k = 1,..., M. The classifier 2 when applied to the image Abb [i (l)] then provides the actual read result Is [i (l), i, k]. If Ist [i (l), i, k] deviates from Ist [i (j)] for at least one image Abb [i (l)], then P (i) is an influencing parameter (result E1 in 9 ). If, on the other hand, the comparison yields the result Is [i (j), 0] = Ist [i (j), 1] = Ist [i (j), m] for each misclassified image, then P (i) is not an influencing one Parameter (result E0). In total, this determines n1 influencing parameters P [i (1)], P [i (2)],..., P [i (n1)].

As just described and through 4 illustrates, in one embodiment, the parameterization by a central computer and p satellite computer is performed. These computers are connected in a local network. The k-th satellite calculator evaluates N1 (k) selected images (k = 1, ..., p) and varies all influencing parameters. The p satellites work in parallel, because nobody needs a result from another satellite computer. Each satellite computer transmits to the central computer, which has found influencing parameters of these satellite computers, so in each case a lot of influencing parameters. Preferably, this set is an index set, ie a subset of {1, 2, ..., n}. The central computer uses each parameter as influencing that has recognized at least one satellite computer as influencing and transmitted to the central computer. The central computer thus forms the union ("union") of all sets of influencing parameters, which are recognized by the satellite computers as influencing.

For every influencing parameter P [i (j)] (j = 1, ..., n1), all N images of the random sample are now examined. Those images are determined for which the comparison at different values for P (i) is different. This determination illustrates 10 ,

• – dem Ist-Leseergebnis, das der parametrierte Klassifikator 2 liefert, und
• – dem abgespeicherte Soll-Leseergebnis

für verschiedene Werte des Parameters P(i) unterschiedlich ausfällt. Diese Abbilder-Menge T(i) fungiert im Ausführungsbeispiel als die Menge der vom Parameter beeinflussten Sensorwerte-Mengen. Vorzugsweise wird jede Abbilder-Menge T(i) in Form einer Indexmenge erzeugt und abgespeichert, also als Teilmenge der Menge {1, 2, ..., N}.As a result, in each case a set T (i) of influenced images is determined for each influencing parameter P (i) in the first step. The image set T (i) consists of the images of those N mailings of the sample for which the result of the comparison between

• - the actual read result, which is the parameterized classifier 2 delivers, and
• - the stored target reading result

varies differently for different values of the parameter P (i). This image set T (i) functions in the exemplary embodiment as the set of sensor value quantities influenced by the parameter. Preferably, each image set T (i) is generated and stored in the form of an index set, ie as a subset of the set {1, 2,..., N}.

In an embodiment are for each influencing parameter P (i) the sensor value quantities influenced by P (i) from all N images selected from the sample.

In a preferred embodiment, the result of the tracing function is reused to restrict the search space. Thanks to this configuration, both the influencing parameters are determined faster and the influenced sensor value quantities calculated more quickly. As explained above, for each parameter P (i) a quantity V (i) is calculated, namely the quantity of those sensor value quantities (images) of the random sample, for the evaluation of which the classifier is used 2 uses the parameter P (i). Preferably, the sensor value amount used by P (i) is selected from the sensor value amounts of V (i) and not from the entire sample, and therefore the time saving is achieved. For only on the evaluation of these sensor value sets (images) P (i) can have an influence. For the other images, the value of P (i) is not used.

The Procedure is first for the influencing Parameter P (1) explained. For example, the Initial value for parameter P (1) 196. Following m = 4 further values for P (1) are used: 176, 190, 202, 216, ie 196 ± 6 and 196 ± 20. In a preferred Embodiment, the other values are symmetrical about the Initial value.

The following table illustrates the result. Each line represents one mailing. 176 190 196 202 216 Ps (1) right right right right not correct Ps (2) reject reject correct, but incomplete. right right Ps (3) right right right right right Ps (N) not correct not correct not correct not correct not correct

In this example, only the reading results for the postal items Ps (1) and Ps (2) - more precisely: for the images of the postal items Ps (1) and Ps (2) - are influenced by the choice of the value for the parameter P (1) , For the postal items Ps (3), ..., Ps (N) the classifier delivers 2 whereas, for each tested value of P (1), the same comparison result is obtained, and in the row, the same comparison result is obtained each time. Thus, only the images of Ps (1) and Ps (2) are affected by P (1).

This is done for each additional parameter, e.g. B. also for the parameter P (2). The initial value for this parameter P (2) is 65. The following m = 4 further values for P (2) are used: 45, 58, 72, 85, ie 65 ± 7 and 65 ± 20. 45 58 65 72 85 Ps (1) reject right right right right Ps (2) reject reject reject reject reject Ps (3.) right right right right right ... ... Ps (N) reject reject not correct not correct not correct

In In this example, only the reading results for the mailings Ps (1) and Ps (N) of the choice of value for the parameter P (2) influences.

This is done for each influencing parameter. By way of example, the result is explained for the last parameter P (n), which is an influencing parameter. The initial value for P (n) is 118. The following m = 4 further values for P (n) are used: 103, 111, 125, 133, ie 118 ± 7 and 118 ± 15. 103 111 118 125 133 Ps (1) right right right right right Ps (2) reject reject reject right not correct Ps (3) reject right right right right Ps (N) not correct not correct not correct not correct not correct

In In this example, only the reading results for the mailings Ps (2) and Ps (3) of the choice of value for the parameter P (n) influences.

As a result of the first step, a subset T (i) of the random sample consisting of the N (i) influenced images (i = 1,..., N) is determined for each parameter P (i). The image set T (i) consists of N (i) data sets with N (i) images of mailpieces - preferably in the form of an index set, ie a subset of {1, 2, ..., N}. Each record represents one mailpiece and includes the m + 1 actual read results that the classifier 2 for the m + 1 values, which were given by way of example for the parameter P (i). For these N (i) <= N postal consignments of the sample, the reading result is influenced by the choice of the value for the parameter P (i), but not for the remaining N - N (i) postal items.

This calculation of the influencing sensor value quantities (images) can also be performed in parallel on the p satellites computers which are in 4 to be shown. The central computer transmits to each satellite computer the same computer-accessible listing of influencing parameters, eg. As an index set of {1, 2, ..., n}. The k-th satellite calculator evaluates N (k) images, among them N1 (k) selected images (k = 1, ..., p). The k-th satellite computer determines from among these N (k) and N1 (k) images for each parameter P (i) the subset T (i, k) of the N (k) images influenced by the parameter T (i) , These k subsets are transmitted to the central computer, and the central computer forms the affected subset T (i) as a union of the k influenced subset T (i, k).

In the above example, the first step gives the following results for the influencing parameters P (1), P (2) and P (n): N (1) = N (2) = N (n) = 2 and T (1): 176 190 196 202 216 Ps (1) right right right right not correct Ps (2) reject reject reject right right T (2): 45 58 65 72 85 Ps (2) reject reject reject right not correct Ps (N) reject reject not correct not correct not correct T (n): 103 111 118 125 133 Ps (2) reject reject reject right not correct Ps (3) reject right right right right

in the Embodiment will be the second step for every parameter P (i) for which the Number N (i)> 0, so for each parameter, the evaluation at least an image of the sample. If for one Parameter P (i) is the number N (i) = 0, so if P (i) is on does not affect an image of a mail item, so in the second step the initial value Ini (i) is used for P (i) and not anymore changed.

In the second step, each parameter P (i) is optimized locally, for which N (i)> 0 (i = 1, ..., n). The second step provides for each influencing parameter P [i (j)] a local optimum Opt_l [i (j)] (j = 1,..., N1). 11 illustrates how the local optimum of an influencing parameter P [i (j)] is calculated.

Around to locally optimize an influencing parameter P [i (j)], for which N [i (j)]> 0, is given to each of the remaining n - 1 parameters as a value one value each from the value range of the respective parameter assigned. Preferably, every other parameter P (l) the respectively given initial value Ini (l) is assigned (l = 1,. n; l # i [j]).

A local optimization is performed, in which in one embodiment only the parameter P [i (j)] is variable. The parameter P [i (j)] is optimized with respect to the predetermined weighting function Bew, by a local optimizer 11 , Preferably, for local optimization, the set T (j) of the images affected by the parameter P [i (j)] is used as the sample with which the parameter P [i (j)] is locally optimized. For the parameter P [i (j)], a value is found which, in conjunction with the values of the remaining n-1 parameters, leads to a parameterization of the classifier 2 leads to the greatest possible (or small) value of the weighting function Bew. The local optimization supplies a local optimum Opt_l [i (j)] for each influencing parameter P [i (j)].

The n1 influencing parameters can be optimized independently of each other. The n1 local optimizations can therefore be divided up to n1 different computers, z. B. on the p satellite computers that are in 4 to be shown.

The second step is first explained using the example of parameter P (1). The set T (1) is: T (1): 176 190 196 202 216 Ps (1) right right right right not correct Ps (2) reject reject reject right right

A step size S1 (1), z. S1 (1) = 1. The value for P (1) is incremented by S1 (1) repeatedly beginning at 176. The remaining parameter values remain unchanged. The classifier 2 is parameterized on a trial basis, whereby the parameter P (1) is assigned the incrementally incremented value and all other parameters are assigned the initial value. For each value combination the classifier becomes 2 reapplied by the classifier 2 by using the algorithm, parameterized with this parameter set, which automatically evaluates images. The classifier 2 delivered reading result is compared with the target reading result. The desired reading result is stored in the data record for the mailpiece. The default weighting function Bew is used as the target function to locally optimize the parameter P (1). For example, Opt_l (1) = 198 is found as the local optimum.

The local optimum for P (2) obtained by subset T (2) which is affected by P (2) affected images is z. B. 55. The local optimum for P (3), by means of the image set T (3) is found, z. 123.

As stated above, the classifier 2 in one embodiment, properly read the delivery addresses on N2 mailpieces. In one embodiment, this requirement is taken into account as a boundary condition which must be observed in local optimization. In another preferred embodiment, this requirement is not met in the second step, but only in the subsequent third step.

In one embodiment, the local optima of the n1 influencing parameters are determined independently. In this case, n1 optimizations are performed by one function each with one variable leads. The remaining parameters remain constant during such local optimization. These local optimizations can be distributed to different computers, eg. B. on the p satellite computers.

In another embodiment, the classification performance is improved by searching for parameter clusters of strongly correlated influencing parameters. Although the strongly correlated parameters of such a parameter cluster can be assigned values independently, they can be independently optimized locally. However, the classification performance of the classifier becomes 2 then improved if the parameters of a parameter cluster are changed in consideration of a functional relationship between them. This relationship need not be known and it usually is not.

One such parameter cluster of several correlated parameters is characterized by the fact that the parameters of this cluster are strong correlated and therefore essentially the same images of the sample influence.

As outlined above, is the first step for each influencing Parameter P (i) a subset T (i) of those images of the sample, which are influenced by P (i) determined. T (i) consists of N (i) Images. Preferably only parameter clusters of each determines two influencing parameters; is possible but also a parameter cluster with more than two parameters.

In order to determine parameter clusters with two influencing and correlated parameters, a correlation measure Korr (i, j) is calculated for each pair P (i), P (j) of influencing parameters. This correlation measure is preferably between 0 and 1 (incl.) And is calculated according to the following calculation rule:

in this connection card (A) denotes the number of elements of a set A, A ∩ B is the intersection and A ∪ B the union (union) of two sets A and B. The correlation measure Korr (i, j) is equal to 1 when P (i) and P (j) influence the same sets T (i) and T (j), respectively So T (i) = T (j) influence.

For the n1 influencing parameters are thus ( n1 2 ) = n1 * (n1 - 1) / 2 Correlation measures calculated.

A Correlation threshold Korr_min is specified, z. B. 0.8. Everyone affecting parameters belongs to none or exactly one Parameters cluster. The two parameters with the highest Correlation measure the first parameter cluster. Under the remaining n1 - 2 influencing parameters form the both parameters with the second highest correlation measure the second parameter cluster and so on until the remaining parameters all with correlation measures below the correlation threshold Correl_min are correlated.

For each parameter cluster, a local optimization is performed. Here, the two influencing parameters P (i), P (j) of the parameter cluster are variable during local optimization, and each other parameter P (k) retains its initial value Ini (k). The local optimization finds an optimal value Opt_l (i) or Opt_l (j) for P (i) and P (j). At local optimization, the classifier becomes 2 preferably applied to the set T (i) ∪ T (j), that is, to the set of those images of the sample that are influenced by P (i) or by P (j) or by both parameters.

Of the second step thus provides for each influencing parameter P [i (j)] in each case a local optimum Opt_l [i (j)], which only by means of those mailing images were found whose reading result is influenced by this parameter at all.

In a third step, a global optimum for the entire set of n parameters is determined, namely by a global optimizer 12 , 12 illustrates the global optimization. In the embodiment with a central computer and p satellite computers, the central computer preferably carries out the global optimization.

The global optimizer 12 performs a multi-dimensional optimization process, for which the global optimizer 12 the predetermined evaluation function Bew is again used as the objective function to be optimized. The classifier is repeated 2 - parameterized by a parameter value set - on the figure of the N mail items of the entire sample applied, and the reading results of the classifier 2 returns are compared with the stored desired reading results. Although it would also be possible to use the classifier 2 apply only to the affected images (on T [i (1)] ∪ T [i (2)] ∪ ... ∪ T [i (n1)]). However, the classifier should 2 Of course, the unaffected images optimally classify, so preferably the entire sample is used.

at The global optimization will only affect the n1 parameters varies, ie those parameters for which in the first Step one non-empty subset of affected images and in the second step, a local optimum was calculated.

To The beginning of the global optimization of the third step will be everyone Parameter assigned a start value. As starting value for a non-influencing parameter is preferably the initial value used.

Of the Start value for an influencing parameter depends in the general case, from the given initial value or from the local one Optimum, which was calculated in the second step. In a Embodiment is used as starting value for each parameter P (i) whose initial value Ini (i) is used. For global optimization each parameter is only varied within its value range.

Because in the third step only the n1 influencing parameters are varied, in the global optimization much fewer parameters have to be varied than if the parameterization unit 4 Without performing the first and second steps, the same would have done the global optimization and varied all the parameters.

In one embodiment, in the third step the trial parameterized classifier 2 applied only to those images that are affected by at least one influencing parameter. No parameters affect the other images, so these other images are ignored in the third step. This embodiment saves computing time in the third step.

In an alternative embodiment, the trial parameterized classifier 2 in the third step apply to all N images Abb (1), ..., Abb (N). This will ensure that the third step becomes a classifier 2 which works at least as well (in relation to the weighting function Bew) as the classifier parameterized with the initial values 2 ,

In global optimization, a stepwise optimization method is preferably used, e.g. B. a method of steepest rise. For this purpose, all influencing parameters are changed to their respective local optimum. This particular direction to the local optimum provides a "useful direction" for the ramp-up process, e.g. In "Dubbel - Paperback for Mechanical Engineering", 18th ed., Springer-Verlag, 1994, A120-A121, outlined.

In an embodiment that saves computing time, are for each influencing parameter P (i) x values w [i, 1], ..., w [i, x] from the value range W [i] of the influencing parameter P (i) selected. Preferably, w [i, 1] ≔ Ini (i) and w [i, x] ≔ Opt_l (i). If the initial value Ini (i) is smaller than the local optimum Opt_l (i), then the values become chosen so that w [i, 1] <w [i, 2] <... <w [i, x]. in the otherwise, they are chosen such that w [i, 1]> w [i, 2]> ...> w [i, x]. If Ini (i) = Opt_l (i), this parameter P (i) becomes the third step does not vary but retains its value Ini (i) = Opt_l (i).

The classifier 2 is applied x times to the N images of the sample, preferably to the entire sample with the N images. In the first classification run, each influencing parameter P [i (j)] is assigned the value w [i (j), 1] (j = 1, 2,..., N1) and every other parameter has its respective initial value. The classification performance of the classifier parameterized in this way 2 in the first classification run, the given evaluation function Bew is evaluated. In the second classification run, each influencing parameter P [i (j)] is assigned the value w [i (j), 2] (j = 1, 2,..., N1) and every other parameter has its respective initial value. This is done accordingly for the subsequent classification runs.

Subsequently, the classification achievements achieved in this way are compared. The best classification performance is achieved by the classifier 2 in the classification run x1 (1 <= x1 <= x), in which each influencing parameter P [i (j)] has been assigned the value w [i (j), x1]. As the global optimum Opt [i (j)], the influencing parameter P [i (j)] is assigned the value w [i (j), x1].

The x classification runs of the classifier 2 with the x different parameter value sets can be carried out in parallel and distributed on up to x computers. In the design with the central computer and the p satellite computers, the x classification runs are distributed to the p satellite computers.

The Optimization procedure of the third step is terminated if at least a predetermined abort criterion is met, z. For example a predetermined number of iterations has been performed or the goal function values for the last r examined Parameter value sets less than a predetermined threshold differ. Other possible termination criteria are that the term, the implementation of the global optimization so far has reached a predetermined time limit or that in the third step a predetermined number of iterations was carried out. It is also possible that the third step is then canceled if at least one of several predetermined termination criteria is met.

As stated above, the classifier 2 in one embodiment, properly read N2 mailpieces. This requirement is used in the global optimization of the third step as a constraint that needs to be met.

After the third step, each parameter is assigned a value. In the exemplary embodiment, each non-influencing parameter retains its initial value. Each influencing parameter was given different values on a trial basis and finally receives a value which - together with the other parameter values - results in a well-parameterized classifier 2 leads. Of course, it is possible that an influencing parameter afterwards has the initial value again.

One but gradual optimization process provides in the not necessarily an optimum. Therefore, will In one embodiment, the parameterization in several parameterization runs carried out.

In front Each parameterization run is assigned a start value for each parameter assigned. During each parameterization run, the three steps described above performs what the N images Abb (1), ..., Abb (N) of the sample are used. Each parameterization run returns one value for each influencing parameters. It is possible that a parameter is an influencing parameter in a parameterization run and not in the next, because it is different from the values of others Parameter may depend on whether a parameter is influencing Parameter is determined or not.

At the The first parameterization run will be the default initial values Ini (1), ..., Ini (n) used as starting values of the parameters. at for each additional parameterization run, the starting values are those Parameter values Opt (1), ..., Opt (n) used in the preceding Parameterization run were calculated as the global optimum.

• – Die Bewertungsfunktion Bew, die der parametrierte Klassifikator 2 nach einem Parametrierungslauf erreicht, unterscheidet sich nur noch um weniger als eine vorgegebene Verbesserungs-Schranke von der aus dem vorhergehenden Parametrierungslauf.
• – Eine vorgegebene Maximalanzahl von Parametrierungsläufen ist erreicht.
• – Die Parametrierung hat eine vorgegebene Zeitspanne an Rechenzeit verbraucht.

The execution of parameterization runs is aborted at the latest when the first step of a parameterization run yields the result that no parameter is an influencing parameter. Further termination criteria can be:

• - The valuation function Bew, which is the parameterized classifier 2 After reaching a parameterization run, the difference is only less than a predetermined improvement threshold from that from the previous parameterization run.
• - A predefined maximum number of parameterization runs has been reached.
• - The parameterization has consumed a specified amount of computing time.

In one embodiment, it is determined what effect the constraint that the N2 delivery addresses are to be read correctly on the performance of the classifier 2 Has. As a test, at least one of the N2 images is selected and removed from the constraint to be met. The third step, the global optimization, will be carried out again. Here, however, only the weakened boundary condition is used, that the classifier 2 must properly read the delivery addresses on the remaining unselected N2 images. The reading power with the weakened boundary condition is compared with the reading power that causes the original boundary condition.

In a modification, again at least one of the N2 images is selected. This image must be classified correctly. Furthermore, an influencing parameter P [i (j)] is selected such that the selected image belongs to the images influenced by the selected influencing parameter. If the global optimization is carried out differently, the selected influencing parameter P [i (j)] is no longer varied but retains the value Opt_l [i (j)] from the second step or also at its In Ini (i), or more generally, to a value that properly classifies the selected image. This can ensure that a particular image is properly classified, thereby satisfying a constraint. This can also be done for several parameters.

In another boundary condition, it is measured in advance how much time the classifier is 2 averaged to classify one of the N images of the sample, if the classifier 2 is parameterized with the initial values. The third step is specified as a boundary condition that the classifier 2 then, when parameterized with the results of global optimization, may not require more than x times the pre-measured run time per image, where x is a predetermined factor, e.g. B. x = 3rd list of reference numerals reference numeral importance 1 sorting 2 Classifier of the sorting plant 1 2.1 . 2.2 , ... Single classifiers of the classifier 2 3 Camera of the sorting plant 1 , the image produces Fig 4 parameterization unit 5 Database of the classifier 2 6 Calculator of the classifier 2 6.1 . 6.2 Two satellite computers used in the parameterization phase. 7 Another computer, the parameterization of the classifier 2 performs 8th Sampling data storage 11 local optimizer 12 global optimizer 13 Parameter values assignment unit α Weighting factor for Ri in the weighting function Bew α (l) Weight factor for the address component type l (l = 1, ..., s) Fig computer-accessible image of the postal item Ps Fig (1), ..., Fig (N) Sample with N sensor value quantities (images of postal items) Af-A, Af-B, ..., Af-E Output compartments of the sorting system 1 Al.1, Al.2, ... Algorithms, which are the single classifiers 2.1 . 2.2 , ... apply β Weighting factor for NL in the evaluation function Bew Bew default valuation function Card (A) Number of elements of the set A E.1, E.2, ... Results of the single classifiers 2.1 . 2.2 , ... γ (k) predefined weight factor for a reading error type in the evaluation function Ini (i) given initial value for the parameter P (i) (i = 1, ..., n) Is [l] Actual read result, which is the classifier 2 for the image fig (l) then returns, if the classifier 2 is parameterized with the initial values (l = 1, ..., N) Is [l, i, k] Actual read result, which is the classifier 2 for the image Fig (l) then yields, if the parameter P (i) has the value w [i, k] and otherwise the classifier 2 is parameterized with the initial values (i = 1, ..., n and l = 1, ..., N and k = 1, ..., m) K (i) Identifier with which the postal item Ps (i) of the random sample is provided (i = 1,..., N) Corr (i, j) Correlation measure between the influencing parameters P (i) and P (j) LE Reading result of the classifier 2 namely, the delivery address XYZ LF (k) Percentage of read errors of type k on the total read strings (k = 1, ..., r) n Number of parameters of the algorithm Al n1 Number of influencing parameters, for each of which a value is determined N Number of mailpieces in the sample N1 Number of selected postal consignments, that is to say those postal consignments of the random sample in which the classifier parameterized with the initial values 2 returns a wrong or incomplete reading result N2 Number of those mailings in the sample whose string must be read correctly NL Percentage of strings classified as unreadable Opt [i (j)] globally optimal value for the influencing parameter P [i (j)], which is calculated in the third step (j = 1,..., n1) Opt_l [i (j)] locally optimum value for the influencing parameter P [i (j)], which is calculated in the second step (j = 1,..., n1) p Number of satellite computers of the local network P (1), ..., P (n) Parameters of the algorithm Al and thus of the classifier 2 , for each of which a value is determined P [i (1), P [i (2)], ..., P [i (n1)] In the first step determined influencing parameters ps Mailing with the delivery address XYZ coming from the sorting system 1 in one of the output compartments Af-A, Af-B, ..., Af-E is discharged. Ps (1), ..., Ps (N) Mail items in the sample r Number of types of read errors that are distinguished Ri Percentage of correctly read strings Ri (l) Percentage of correctly read address components of the type l (l = 1, ..., s) s Number of address components S1 Preparatory step: N images Abb (1), ..., Abb (N) of the N mail items Ps (1), ..., Ps (N) generate and save S2 Preparatory step: Determining and storing desired read results of the N mail items Ps (1),..., Ps (N) T (j) Subset of those images of the sample for which the reading result is affected by the value for the parameter P [i (j)] (j = 1, ..., n1) tab Table with the desired reading results X [1], X [2], ..., X [N] for the N mailings of the sample V (i) Quantity of those images (sensor value sets), for their evaluation the classifier 2 the parameter P (i) is used when the classifier 2 is parameterized with the initial values W [i] Value range of the parameter P (i) (i = 1,..., N) w [i, k] Value from the value range W [i], which is assigned to the parameter P (i) on a trial basis (i = 1, ..., n and k = 1, ..., m) x Number of classification runs in the global optimization in the third step X [i] Delivery address with which the postal item Ps (i) of the random sample is provided (i = 1, ..., N) XYZ Delivery address of the mail item Ps

QUOTES INCLUDE IN THE DESCRIPTION

This list The documents listed by the applicant have been automated generated and is solely for better information recorded by the reader. The list is not part of the German Patent or utility model application. The DPMA takes over no liability for any errors or omissions.

Cited patent literature

• - US 7031512 B2 [0003]
• US 7031519 B2 [0023]

Cited non-patent literature

• - A. Berlanga et al .: "Optimizing Statistical Character Recognition Using Evolutionary Strategies to Recognize Aircraft Tail Numbers", EURASIP Journal on Applied Signal Processing 2004: 8, pp. 1125-1134 [0005]
• - "Dubbel - paperback for mechanical engineering", 18th ed., Springer-Verlag, 1994, A120-A121, outlined. [0176]

Claims (26)

A method for automatically classifying a physical object (Ps), the classifying comprising the steps of - a sensor ( 3 ) Taking measurements on the physical object (Ps), - the sensor value set (Fig) generated by the measurement is stored in computer-available form, - a classifier ( 2 ) classifies the stored sensor value set (Fig) by applying an algorithm (Al) and - the classifier ( 2 ) as a classification of the object (Ps) either the result of the evaluation of the sensor value set (Fig) or the result, the sensor value set (Fig) can not classify, where - the algorithm (Al) of several parameters (P (1), ..., P (n)) whose values remain unchanged during classifying, and - the classifier ( 2 ) is automatically parameterized before classification, whereby each parameter (P (1), ..., P (n)) of the algorithm (Al) is assigned an automatically determined value, for the parameterization - an object sample with N physical Sample objects (Ps (1), ..., Ps (N)), - a sensor ( 9 ) generates a sample of N sensor value quantities (Fig. (1),..., Fig. (N)) by measurements on the N sample objects (Ps (1),..., Ps (N)), the be stored in computer-available form, - for each sensor value set (Fig (1), ..., Fig (N)) of the sample a desired classification result (X [1], ..., X [N]) determined and is stored in computer-available form, - a rating function (Bew) for evaluating a parameterization is given, which depends on the rate of incorrectly classified sensor value quantities and the rate of the sensor value sets classified as non-classifiable, parameterizing the classifier ( 2 ) comprises a first step, a second step and a third step, in each step - the classifier ( 2 ) is repeatedly parameterized with different parameter values, - the parameterized classifier ( 2 ) is applied to at least one sensor value set of the sample (Fig. (1), ..., Fig. (N)) and - for each applied sensor value set (Fig. (1), ..., Fig. (N)) the respective actual classification result that the classifier ( 2 ) is compared with the stored desired classification result (X [1], ..., X [N]), in the first step at least one influencing parameter (P [i (1)]), ..., P ( [i (n1)]) of the algorithm (Al), which influences at least one sensor value set of the sample (fig. (1),..., fig. (n)), one parameter then being a sensor value set Sample (Fig (1), ..., Fig (N)), if the result of the comparison between - the actual classification result, the classifier ( 2 ) applied to this sensor value set, and - the stored desired classification result (X [1], ..., X [N]) for this sensor value set (Fig. (1), ..., Fig. (N) ) is different for different values of the parameter, in the second step for each influencing parameter (P [i (1)]),..., P ([i (n1)]) determined in the first step a local optimum (Opt_l [] i (1)]), ..., Opt_l ([i (n1)]) is calculated, wherein in the second step in the local optimization of an influencing parameter - the predetermined evaluation function (Bew) is used as the target function to be locally optimized, and - the classifier ( 2 ) is parameterized with different values for the influencing parameter, in the third step a global optimum for the parameters (P (1), ..., P (n)) is calculated, wherein in the third step - the predetermined evaluation function (Bew) is used as the global objective function to be optimized, - each influencing parameter is varied using the local optimum calculated in the second step, and the classifier ( 2 ) is parameterized for classifying the object with those parameter values which were calculated in the third step in the global optimization. Method according to claim 1, characterized in that the classifier ( 2 ) is parameterized with one initial value (Ini (1), ..., Ini (n)) for each parameter (P (1), ..., P (n)) before the first step, the classifier parameterized with the initial values (Ini (1), ..., Ini (n)) ( 2 ) is applied to each sensor value set of the sample (Fig (1), ..., Fig (N)), the respective actual classification result that the classifier ( 2 ) for this sensor value quantity, which is compared with the stored desired classification result of this sensor value set, as incorrectly classified sensor value quantities those sensor value quantities of the random sample (FIGS. (1),..., fig. (N)) where this actual classification result deviates from the desired classification result, and in the first step all those parameters of the algorithm (A1) which influence at least one wrongly classified sensor value set are determined, and the parameters determined in this way are the influencing parameters (P [ i (1)]), ..., P ([i (n1)]). Method according to claim 1 or claim 2, characterized in that, before determining the influencing parameters, it is determined for each parameter (P (1), ..., P (n)) of the algorithm (A1) whether the classifier ( 2 ) This parameter is used to classify at least one sensor value set of the sample (Fig. (1),..., Fig. (N)) or not, and the influencing parameters are determined exclusively under the parameters used. Method according to Claim 3, characterized in that for each parameter (P (1), ..., P (n)) of the algorithm (A1) it is determined which sensor value quantities of the random sample (Fig. (1), ... , Fig (N)) the classifier ( 2 ) is classified using this parameter, and the sensor value quantity (T (i)) influenced by this parameter is determined among the sensor value quantities which the classifier ( 2 ) using this parameter. Method according to one of claims 1 to 3, characterized in that the sample (Fig (1), ..., Fig (N)) is divided into at least two partial samples, For Each sub-sample is determined which parameters of the algorithm (Al) at least one sensor value set of the sub-sample influenced, and as influencing parameters those parameters At least one sensor value set should be used influence a partial sample. Method according to one of claims 1 to 5, characterized in that in the first step for each influencing parameter (P [i (1)]), ..., P ([i (n1)]) all sensor values influenced by this parameter Quantities of the sample (Fig (1), ..., Fig (N)) are determined and in the second step in the local optimization for an influencing parameter - the classifier ( 2 ) is repeatedly parameterized with a value for this influencing parameter and - the classifier parameterized in this way ( 2 ) is applied to all sensor value quantities affected by this parameter. Method according to claim 6, characterized, that the sample (fig (1), ..., fig (N)) in at least two Sub-samples is subdivided, in the first step for influenced each influencing parameter and each sub-sample all Sensor values quantities of the partial sample are determined and when affected sensor value set the sample the union amount all thus determined influenced sensor value amounts used becomes. A method according to claim 6 or claim 7, characterized in that in the second step in the local optimization of an influencing parameter of the sample-parameterized classifier ( 2 ) is applied exclusively to the sensor value quantities influenced by the parameter. Method according to one of claims 1 to 8, characterized in that in the second step, the local optimization temporally overlapping by at least two influencing parameters is carried out. Method according to one of claims 1 to 9, characterized in that in the second step in the local optimization of an influencing parameter of the classifier ( 2 ) is parameterized on a trial basis in such a way that the values of all other parameters remain unchanged during this local optimization. Method according to one of claims 1 to 10, characterized in that at least two parameters are selected from the influencing parameters and combined into a parameter cluster, wherein the selection of the at least one parameter cluster is performed such that each influencing parameter to at most one Parameter cluster belongs and each parameter cluster consists of fewer parameters than total parameters to be determined in the first step, and in the second step the parameters of each parameter cluster are optimized in such a way that the classifier ( 2 ) is tentatively parameterized with different values for the parameters of the parameter cluster and the values of all parameters not belonging to the parameter cluster remain unchanged. Method according to one of Claims 1 to 11, characterized in that in the third step the test parameterized classifier ( 2 ) is applied only to those sensor value quantities that are affected by at least one parameter. Method according to one of Claims 1 to 11, characterized in that in the third step the test parameterized classifier ( 2 ) is applied to all sensor value quantities of the sample (Fig. (1),..., Fig. (N)). Method according to one of claims 1 to 13 characterized in that in the third step - each Parameter (P (1), ..., P (n)) as the starting value in each case a predetermined Initial value (Ini (1), ..., Ini (n)) is assigned and - then, if the local optimum of an influencing Parameter differs from the assigned initial value, - the influencing parameters at least once a test value is assigned which differs less from the local optimum as the initial value. Method according to one of claims 1 to 14, characterized in that the classifier ( 2 ) is parameterized with a given initial value for each parameter and, in the third step, for each parameter which does not influence a sensor value set of the sample, - the assigned initial value is used as initial value and - the value of this parameter remains unchanged. Method according to one of claims 1 to 15 characterized in that for at least a sensor value set of the sample next to the target classification result In addition, at least one correct, but incomplete Classification result is determined and stored and the default valuation function (Bew) additionally from the Quota of correctly, but incompletely classified sensor value quantities depends. Method according to one of claims 1 to 16, characterized in that - at least one object of the N sample objects (Ps (1), ..., Ps (N)) is given as an object to be classified correctly, - in the third step global optimization is performed under the constraint that the classifier ( 2 ) is to be parameterised so that the classifier ( 2 ) provides for each sensor value set of an object to be correctly classified an actual classification result that matches a stored target classification result for the object. Method according to one of claims 1 to 17, characterized in that each parameter of the algorithm is assigned a predetermined initial value before the first step and then, if no influencing parameter is determined in the first step, - the second and the third step are omitted and - the classifier parameterized with the initial values ( 2 ) classifies the object. Method according to one of claims 1 to 18, characterized in that the parameterization of the classifier ( 2 ) is carried out in several parameterization runs, wherein each parameter is assigned a predetermined initial value before starting the first parameterization run, the first step, the second step and the third step are performed in each parameterization run and before each parameterization run each parameter the start value is assigned to the value determined by the third step of the previous parameterization run. Method according to one of claims 1 to 19, characterized in that the classifier ( 2 ) additionally classifies a safety measure when classifying the object, the classifier ( 2 ) then yields the result of being unable to classify the object if the security measure is below a security threshold, the security threshold being a parameter of the algorithm for which a value is determined during parameterization. Method according to one of claims 1 to 20 characterized in that any amount of sensor values is generated thereby that an object is optically scanned becomes. Method according to claim 21, characterized, that - Each object scanned optically with one String is provided, - the optical scanning the item comprises the step of having an image of the string generated and used as a sensor value set, - the Classifying each sensor value amount includes the step through Evaluating the string image the string automatically to read, and - as classification result a coding the respectively read string is used. Method according to one of claims 1 to 22 characterized in that any amount of sensor values is generated thereby that a sequence of sounds is measured. Method according to claim 23, characterized, that - every noise measured a spoken one Word includes, - classifying each sensor value set includes the step that by the evaluation of the measured Sound that at least one contained in the sound spoken word is automatically recognized and - when Classification result a coding of the automatically recognized at least one word is used. Disk with a computer program, the computer program being designed to: - be loaded into a memory of a data processing system, - to be executed by the data processing system and - to execute a method according to one of claims 1 to 24 when executed. Device for automatically classifying a physical object (Ps), the device comprising - a sensor ( 3 ), - a classifier ( 2 ), - a parameterization unit ( 4 ) and - a sample data store ( 8th ), the sensor ( 3 ) is configured to - take measurements on a physical object (Ps) and - store the sensor value set (fig) generated by the measurement in computer-available form, the classifier ( 2 ) is designed to - evaluate the stored sensor value set (Fig) by using an algorithm (Al) and - as a classification of the object (Ps) either the result of the evaluation of the sensor value set (Fig) or the result, the sensor values- The algorithm (Al) depends on several parameters (P (1), ..., P (n)) whose values remain unchanged during classification, the parameterization unit ( 4 ) is designed to - the classifier ( 2 ) to be parameterized automatically before classifying, whereby the parameterization unit ( 4 ) assigns to each parameter (P (1), ..., P (n)) of the algorithm (A1) an automatically determined value, and - to apply a predetermined evaluation function (Bew) to the parameterization, which depends on the rate of the incorrectly classified sensor values Quantities and the rate of sensor value quantities classified as non-classifiable, in the sample data memory ( 8th ) - a sample of N sensor value quantities (Fig. (1),..., Fig. (N)) is stored or storable, which is determined by measurements on N predetermined physical sample objects (Ps (1),..., Ps (N)) is generated, and - a set classification result (X [1], ..., X [N]) is stored for each sensor value set (Fig (1), ..., Fig (N)) or can be stored, the parameterization unit ( 4 ) is configured to parameterize the classifier ( 2 ) in a first step, a second step and a third step in such a way that the parameterization unit ( 4 ) in each step - the classifier ( 2 ) repeatedly parameterized with different parameter values, - the parameterized classifier ( 2 ) applies to at least one sensor value set of the sample (Fig. (1),..., Fig. (N)) and - for each of these sets of sensor values (Fig. (1),..., Fig. (N)) Actual classification result that the classifier ( 2 ), compares with the stored desired classification result (X [1], ..., X [N]), the parameterization unit ( 4 ) in the first step, at least one influencing parameter of the algorithm (Al), which influences at least one sensor value set of the sample (fig (1), ..., fig (N)), wherein a parameter then a sensor value set the Sample (Fig (1), ..., Fig (N)), if the result of the comparison between - the actual classification result, the classifier ( 2 ) applied to this sensor value quantity, and - the stored nominal classification result for this sensor value set differs for different values of the parameter, the parameterization unit ( 4 ) in each case a local optimum is calculated for each influencing parameter determined in the first step, wherein the parameterization unit ( 4 ) in the second step in the local optimization of an influencing parameter - the predetermined evaluation function (Bew) is used as the target function to be locally optimized and - the classifier ( 2 ) is parameterized on a trial basis with different values for the influencing parameter, the parameterization unit ( 4 ) in the third step calculates a global optimum for the parameters, where the parameterization unit ( 4 ) in the third step - the predefined evaluation function (Bew) is used as the target function to be optimized globally, - each influencing parameter is varied using the local optimum calculated in the second step and - the classifier ( 2 ) for classifying the object with the parameter values calculated in the third step in the global optimization.