Method and device for evaluating valuable documents

Abstract

The invention relates to a method for generating reduced-distortion training data in order to determine parameters of a test model for testing valuable documents of a specified valuable document type, in particular banknotes. The method comprises acquiring a plurality of spatially resolved raw data for each training valuable document of a number of training valuable documents by means of at least one sensor device. The raw data comprise sensor values. The method comprises ascertaining at least one compensation value for the sensor values of the raw data on the basis of a predetermined compensation distribution. The method comprises changing the sensor values of the raw data on the basis of the at least one compensation value in order to generate the reduced-distortion training data.

Description

[0001] Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE

[0002] Method and device for evaluating securities

[0003] The present invention relates to a method for evaluating securities, more precisely for generating distortion-reduced training data for determining parameters and for determining the parameters based on the generated training data. The parameters are for use in evaluating, and in particular verifying, securities. The invention further relates to corresponding devices.

[0004] Securities are understood to be sheet-like objects that represent, for example, a monetary value or an entitlement and are therefore not intended to be arbitrarily produced by unauthorized persons. They therefore possess security features that are not easily manufactured, particularly copied, and whose presence is an indication of authenticity, i.e., production by an authorized entity. Securities can take different forms. Important examples of such securities are coupons, vouchers, checks, and especially banknotes. In the case of banknotes, securities types can be further differentiated; within the scope of the present invention, a securities type can be defined by the currency and face value of a banknote, or its denomination, and optionally by its issuance or the period during which it was officially issued, for example, by central banks.Insofar as the following explanations refer to banknotes, they apply accordingly to any other type of security document.

[0005] In practice, valuable documents, especially banknotes, are evaluated in two ways. Firstly, their authenticity is verified, which includes checking security features that are difficult to copy and whose presence can be considered an indicator of the document's authenticity. Secondly, valuable documents, especially those that have already been in circulation, are examined for their condition: for example, they must still be recognizable as such with sufficient certainty and, in particular, be processable by simple machines, such as ATMs or banknote validators. Condition criteria include, for example, the presence of damage such as tears or missing parts, creases or dog-eared pages, adhesive tape, or the document's flimsiness. Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE

[0006] For the evaluation of valuable documents, document processing devices are used. These devices separate the documents, which are presented in stacks, and then transport them individually past sensors that record the relevant properties of each document. The sensor data is evaluated in real time to verify the respective document. Depending on the verification results, the documents are then sorted into appropriate output areas.

[0007] When examining the condition, or so-called fitness, of a security, it is evaluated whether it meets certain criteria to be put back into circulation or to be removed from circulation. In addition to the degree of soiling and wear, the presence of unwanted foreign objects, usually in the form of adhesive tape or other stickers, on the banknote is also an important criterion.

[0008] The presence of adhesive tape is usually checked by measuring the thickness of the banknote to be checked using mechanical or ultrasonic thickness sensors.

[0009] When evaluating a security document of a predefined document type, sensor data is often acquired from the document being examined. This data corresponds to measurement points arranged in a grid pattern on the document. The sensor data can represent brightness, color, thickness, etc., of the respective measurement point. During the examination, depending on the method used, it is evaluated whether, and to what extent, the sensor data fulfills at least one predefined test criterion.

[0010] The sensors used have multiple sensor elements arranged transversely to the transport direction T of the valuable documents. During transport, each sensor element successively detects a section of the valuable document within its detection range. Therefore, the sensor data is available in a track-by-track fashion, with each track corresponding to a sensor element. The measurement locations along the track depend on the transport speed and the sensor element's measurement rate. The measurement data from the parallel tracks generates a grid of two-dimensional sensor data for areas of the valuable document or the entire document. Magnetic and ultrasonic sensors are among the types used. Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE

[0011] To evaluate the acquired sensor data, the evaluation algorithm requires parameters or test parameters that are typically specific to the type of value document, the type of sensor, and the property being tested. Determining these specific test parameters can be described as adaptation. This is achieved by acquiring sensor data from training value documents using a value document processing machine. Ideally, the sensor used should be calibrated. The parameters for the evaluation algorithms can then be optimized based on the training data. Evaluation algorithms can also be parameterized so that the specified parameters are suitable for evaluating multiple types of value documents.

[0012] The sensors are typically calibrated to ensure that the measurement characteristics of different sensors do not vary. The sensors can also be recalibrated to correct for changes in their measurement characteristics over time. Each calibration generates a set of calibration parameters. These are specific to a particular sensor or sensor element, but are usually independent of the type of document being measured.

[0013] The calibration parameters are determined during the manufacturing of the document processing devices, i.e., initially determined. Therefore, the calibration parameters can be device-specific. For this purpose, the sensors used are calibrated accordingly. Selected sensors, such as ultrasonic sensors used to measure the thickness at the document's measurement points, also recalibrate themselves at regular intervals during operation.

[0014] The parameters of the evaluation algorithm, however, are generally not adjusted during operation. The determined parameters of the evaluation algorithm can be transferred to other document processing devices and used for evaluating documents. Provided the sensor of another document processing machine is calibrated, the evaluation of its sensor data should yield results of similar quality to those of the document processing device with which the parameters were originally determined. The quality of the evaluation results can be characterized, for example, by the number of false alarms during the evaluation of documents. Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE

[0015] It has been found that the testing and sorting performance of document processing systems can deteriorate abruptly during operation. Specifically, it was observed that, with a consistent range of documents, a document processing system can generate a sudden increase in false alarms after conventional adaptation. Conversely, the testing and sorting performance can also improve abruptly after a period of deterioration. A data-driven analysis of the testing behavior revealed a correlation between the regular recalibration of the sensors and the fluctuating quality of the evaluation results.

[0016] The invention therefore aims to provide more robust evaluation algorithm parameters in order to at least partially avoid abrupt deteriorations in the quality of the evaluation results. A further objective of the invention is to determine parameters for a more robust evaluation method for securities based on existing training securities and without increasing the size of conventional training datasets. Another objective of the invention is to provide a method that can evaluate securities with the most consistent predictive quality possible. Finally, the invention aims to provide means for carrying out these methods.

[0017] At least one of these tasks is solved by the features of the independent claims. Further details are specified in the dependent claims.

[0018] In one aspect of the present disclosure, a method for generating distortion-reduced training data for determining parameters of an evaluation procedure for securities of a given security document type, in particular banknotes, is provided.

[0019] The process involves receiving a large number of spatially resolved raw data sets for each training value document within a set of training value documents. The raw data comprises sensor values, each assigned to a location on the training value document. The received raw data can be acquired by at least one calibrated sensor device and constitute a training data set. Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE

[0020] The procedure further includes determining a compensation value for each of the sensor values ​​of at least two disjoint sections of the raw data based on a predetermined compensation distribution.

[0021] The process also includes modifying the sensor values ​​of the raw data based on the respective compensation value to generate the distortion-reduced training data.

[0022] The evaluation procedure is set up to evaluate value documents based on spatially resolved raw data recorded from them, whereby the parameters of the evaluation procedure are determined based on the distortion-reduced training data.

[0023] The method for determining parameters or test parameters can be computer-aided. For this purpose, instructions can be stored in memory within a data processing system comprising a memory and a processor. These instructions can be read, interpreted, and executed by the processor. By executing the instructions, the processor can perform the steps of the method according to the invention. The raw data and / or the generated training data can be stored in the memory. The generated training data can be made available for determining the parameters of a test model in the same data processing system or in a different data processing system. By storing the generated distortion-reduced training data, this data can be made accessible to other data processing systems or computers.

[0024] The inventor recognized that conventionally used training data contains data bias, which impairs the generalizability of test procedures that include parameters generated from this training data. This is because the calibration and recalibration of a sensor are subject to errors. As a result, systematic differences in the measurement behavior of different sensors cannot be completely avoided. Errors during the recalibration of a sensor can lead to differences in the measurement behavior of that individual sensor. Furthermore, in practice, training data is often acquired using only a few sensors and at only a few time points, meaning that the training data does not provide a complete picture of all possible sensor measurement behaviors and can therefore be biased. Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE

[0025] The term "data bias" refers to systematic errors in the raw data that can lead to incorrect conclusions or results. Data bias can occur when certain results are favored over others due to the way data is collected, processed, or analyzed; for example, the measurement behavior of sensors may be underrepresented in the conventional training data after a recalibration error.

[0026] A conventional way to reduce distortions in the training data is undoubtedly to record more training data, particularly in different states of the security document processing device, i.e., after repeated recalibrations of the sensor setup. However, this approach does not solve the problem of the invention. Due to the large number of security document types, increasing the training dataset would entail a significantly higher effort for determining the parameters for the evaluation method and should therefore be avoided.

[0027] A significant advantage of the method according to the invention is that the distortion-reduced training data enables the determination of parameters for a more robust evaluation method without increasing the size of the training dataset. According to the method, the collected raw data are specifically altered or modified, thereby eliminating the systematic error inherent in the recorded raw data due to the calibration status of the document processing device.

[0028] To improve the raw data, a compensation value is first determined for the sensor values ​​of at least two disjoint sections of the raw data. The sensor values ​​of the raw data correspond to the measured values ​​for locations on a grid of a training data document. For example, the sensor values ​​can be positive numbers, particularly integer values ​​representing the thickness or basis weight of the training data document. The compensation value can also be a number, which can be positive or negative. The compensation value can be added to the sensor values. This targeted modification of the received training data set allows for the determination of parameters that ultimately stabilize the quality of the evaluation method. Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE

[0029] The compensation value, i.e., a positive or negative random number by which the sensor values ​​are to be changed, is determined using a compensation distribution. The compensation value can be drawn from a predetermined distribution, preferably statistically independent. The compensation distribution is configured to provide multiple compensation values ​​that are highly likely to lie within a defined interval. This allows the identified systematic error in the collected training data to be at least partially compensated without substantially altering the sensor values. The compensation value can therefore be a random variable distributed according to the compensation distribution, i.e., a predetermined probability distribution.

[0030] The spatially resolved raw data acquired, collected, or recorded from each of the training documents can be organized on a grid representing a (training) document. The measurement locations on the grid can be determined by the position of the sensor array above the training document and a selected sampling rate or transport speed of the (training) documents. Advantageously, the raw data for each training document includes measurement locations on every possible orientation of the document. In the case of a banknote, four orientations can be distinguished: a front side, which can be rotated 180° around a vertical axis, and a back side, which can also be rotated 180° around the same vertical axis.

[0031] The sensor device can comprise two or more sensor elements that simultaneously acquire raw data from different measurement locations on the training value documents. In one embodiment, the raw data comprises sensor values ​​along parallel measurement tracks on each training value document. In other words, the sensor elements can be spaced apart from one another along a direction transverse to the transport direction, so that during a relative displacement between the sensor device and the value document, the sensor elements each scan a path or track on the value document along the transport direction. The individual tracks of the sensor elements on the value document can be disjoint from each other and parallel to each other. The sensor device can, in particular, comprise 4 to 16, preferably 6 to 14, and even more preferably 8 to 12 sensor elements. In one embodiment, the sensor device comprises 10 sensor elements.Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE.

[0032] The altered sensor values ​​can be used as training data to determine the parameters of a computer-aided evaluation procedure. This computer-aided evaluation procedure can, in principle, be based on any classification method. It has been found that determining the parameters based on the altered sensor values ​​significantly improves the quality of the evaluation results compared to recalibrations or self-calibrations of the sensor system.

[0033] Further analyses have shown that an evaluation method comprising parameters determined with sensor values ​​modified according to the invention does not exhibit any abrupt loss of quality in the evaluation results or any deterioration in the classification or prediction rate after self-calibration of the sensor device, without compromising the absolute quality of the evaluation results compared to an evaluation method using conventionally determined parameters. Conventionally determined parameters are those determined based on training data collected directly after perfect calibration of the sensor device, without any adjustment or modification of the training data.

[0034] In other words, an evaluation method whose parameters are determined using modified sensor values ​​can achieve the same absolute metrics as a conventional evaluation method, but without the sensor calibration-related, section-by-section dips or degradation of the metrics. One or more of the following metrics can be selected: accuracy, precision, sensitivity, Fl score, area under the receiver operating curve (ROC-AUC), and / or confusion matrix. The modified sensor values, which significantly influence the performance of the evaluation method due to their contribution to parameter determination, thus produce the extended technical effect of a more robust evaluation result.

[0035] In one implementation, the distortion-reduced training data can additionally include unaltered training data, or the original training data can be supplemented with the distortion-reduced training data. That is, the training data can be enriched so that the parameters can be determined based on the enriched training dataset, which includes, among other things, the distortion-reduced training data. Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE

[0036] In one embodiment, the raw data can assign a sensor value to each recorded location on the value document, whereby changing the sensor values ​​includes adjusting the sensor value, in particular by adding or subtracting at least one compensation value.

[0037] In one embodiment, the disjoint sections of the raw data can comprise at least two tracks, in particular parallel tracks, on the training value documents.

[0038] The procedure may further include determining, for each of the at least two tracks, a compensation value for the sensor values ​​based on the predetermined compensation distribution.

[0039] The procedure can further include modifying, for each of the at least two tracks, the sensor values ​​of the raw data based on the respective compensation value to generate the distortion-reduced training data.

[0040] In this configuration, the sensor values ​​of the raw data can be modified on a track-specific or track-by-track basis; that is, the same compensation value is applied to all sensor values ​​of a track. Therefore, only one compensation value needs to be determined per track.

[0041] These two or more tracks can be detected by moving the valuable document relative to the sensor device or by moving the sensor device relative to the valuable document. The two or more tracks can be detected essentially simultaneously. For this purpose, the sensor device can have two or more sensor elements. The two or more tracks can be detected along a longitudinal direction of the valuable document. For this purpose, the sensor elements in the sensor device can be arranged in a direction perpendicular to the longitudinal direction of the valuable document at a predetermined distance from each other. The predetermined distance determines the width of the respective tracks and can be chosen to be as small as possible to enable essentially complete scanning of the valuable document. The width of the tracks, or track width, defines the resolution of the grid transverse to the transport direction, i.e., in a direction perpendicular to the longitudinal direction of the valuable document.The sampling rate or measurement rate of the sensor elements in conjunction with the speed of the relative movement between Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE.

[0042] The document and sensor setup determine the resolution of the grid along the document's length. Each sensor element can capture raw data along a separate track on the document. These tracks can be arranged non-overlappingly.

[0043] By assigning a compensation value to a track, and thus to a sensor element, the systematic error after self-calibration of the sensor system can be compensated particularly effectively without distorting the sensor values ​​of a sensor element relative to each other. If the sensor system has multiple sensor elements, calibration of the sensor system includes calibration of each individual sensor element. Therefore, after self-calibration of the sensor system, at least one sensor element may have a calibration status that differs from the calibration status of the other sensor elements or that differs from its previous value, thus causing a sudden deterioration in the quality of the evaluation result.Since the systematic source of error was located in the self-calibration of the sensor device, it is advantageous to modify the sensor values ​​collected from the at least two sensor elements on a sensor element-specific basis.

[0044] In one embodiment, changing the sensor values ​​can include adding or subtracting the respective compensation value to or from the sensor values ​​in a track-related manner.

[0045] The sensor values ​​can be changed in such a way that the compensation value determined for each track is added to or subtracted from all sensor values ​​of the respective track.

[0046] In one embodiment, the method can further include acquiring a multitude of spatially resolved raw data for each training value document using at least one sensor device. Preferably, the raw data are acquired by changing a measurement position on the training value document along one of at least two tracks, in particular parallel tracks.

[0047] In one implementation, the predetermined compensation distribution can be a mean-free distribution. The mean-free distribution allows for an additional Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE

[0048] Shifting or distortion of the sensor values ​​averaged over the sensor values ​​of the raw data is avoided.

[0049] In one embodiment, the predetermined compensation distribution can be a normal distribution. The standard deviation of the compensation distribution can be chosen to be on the order of an unavoidable, randomly occurring calibration error. The calibration error typically has a deviation of up to 60 measurement units. The compensation distribution can preferably have a standard deviation in the range of 10 to 100 measurement units of the sensor value, more preferably in the range of 30 to 70 measurement units, and even more preferably in the range of 50 measurement units. The specified ranges for the standard deviation are less than 1% of the typical sensor values ​​for the security document. However, the sensor values ​​of a security document with adhesive tape and one without adhesive tape may also differ only in the single-digit percentage range of the sensor values ​​for the security document.Therefore, the unavoidable, randomly occurring calibration error can have a significant impact on the evaluation of the valuation documents.

[0050] This design of the compensation distribution is advantageous because the normal distribution is a natural assumption for the systematic error resulting from the self-calibration of the sensor system, particularly its individual sensor elements. In other words, it is assumed that a shift in the sensor values ​​of a sensor element compared to the sensor values ​​of the same data set collected before self-calibration is normally distributed. The choice of the standard deviation of the normal distribution allows for the determination of compensation values ​​that lie within the range of the observed shift in sensor values ​​due to self-calibration of the sensor system. Alternatively, instead of assuming a normal distribution, the distribution of the calibration error can also be estimated non-parametrically, for example, using a histogram or a Parzen estimation, also known as kernel density estimation.These configurations of the compensation distribution allow the effect of sensor self-calibration to be artificially and retrospectively applied to the collected raw data without further distorting it. An evaluation method whose parameters are determined using such modified sensor data reacts considerably more robustly to operational self-calibrations of the sensor, thereby improving the stability of evaluations of securities. Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE.

[0051] In one embodiment, the compensation value can represent a random difference between sensor values ​​for the same value document, recorded before and after a self-calibration of at least one sensor device.

[0052] The standard deviation of the compensation distribution used for this purpose is preferably similar in size to the unavoidable sensor calibration error. This refers to a systematic measurement error that remains after calibration. This calibration error can be estimated, for example, by repeatedly measuring an object with one or more sensors, where the desired measured values ​​or sensor values ​​are known, and performing sensor calibrations between measurements. If, for example, it is determined that the unavoidable calibration error can be well described by a normal distribution with a mean of 0 and a standard deviation of 50 units, then it is advantageous to use this or a similar distribution as the compensation distribution.

[0053] In one implementation, the sensor reading can represent the thickness or basis weight of a specific location on the document. A particular pattern of document thicknesses can indicate the presence of foreign objects, such as adhesive tape or stickers.

[0054] In one embodiment, the at least one sensor device can comprise an ultrasonic sensor. The ultrasonic sensor can have at least two ultrasonic transducers or converters, each transducer being assigned to its own track. The ultrasonic sensor can preferably measure ultrasonic transmission at a predetermined ultrasonic frequency. With this configuration, the sensor device can particularly advantageously determine the thickness or basis weight of the document.

[0055] In one embodiment, the at least two tracks can extend parallel and spaced apart along a longitudinal direction of the security document.

[0056] In one configuration, the training data can also include raw data for each training value document in each of its orientations.

[0057] In one configuration, the training value document can be captured once per orientation using the sensor device. In each (main) orientation of the training value document, raw data for measurement locations on a grid of the training value document can be collected or recorded. The main orientations of the value document are the front, in the reading direction and rotated 180° accordingly, and the back, also in the reading direction and rotated 180° accordingly.

[0058] In one embodiment, the (predetermined) number of training documents ranges from 50 to 5000, preferably 100 to 2500, and even more preferably 200 to 1000. With 200 training documents, a training dataset with 800 grids can be obtained as raw data after recording each principal orientation of each training document. Preferably, the evaluation method comprises a number of parameters that is significantly smaller than the number of grids in the training dataset. Such evaluation methods can be designed to be particularly efficient with regard to storage and processing operations and are therefore particularly suitable for implementation in conventional document processing devices. It should be noted that the hardware of document processing devices must be able to operate unchanged for a period of years or even decades, if necessary.by keeping appropriate replacement and spare parts on hand, which can significantly limit computing speeds and capacities.

[0059] The relatively small number of training documents required for reliable parameter determination offers the distinct advantage of time-saving training data creation. A large number of document types are in circulation, each requiring parameter adaptation. Therefore, the time-efficient creation of the respective training data enables a particularly efficient provision of parameters for the evaluation process of a document processing device.

[0060] In a further aspect, a method for determining the parameters of an evaluation procedure for securities of a given security type, in particular banknotes, is provided. The method involves determining the parameters of the evaluation procedure based on distortion-reduced training data generated by a procedure as described above. The method can also be referred to as an adaptation procedure. Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE

[0061] The procedure for determining parameters or test parameters can be computer-aided. For this purpose, instructions can be stored in memory within a data processing system comprising a memory and a processor. These instructions can be read, interpreted, and executed by the processor. By executing these instructions, the processor can perform the parameter determination procedure. The training data can also be stored in memory or provided externally via a data interface.

[0062] For the adaptation process, only training data from securities of the specified security document type is required. Preferably, the training securities can comprise, or preferably consist of, completed and / or newly printed securities of the specified security document type, particularly in the same currency and / or denomination. The training data used should preferably be complete securities, i.e., especially undamaged and undamaged securities.

[0063] Insofar as steps of the methods according to the invention do not use results from other steps, these can be carried out wholly or partially in parallel, nested, or in any order.

[0064] The methods use raw data from securities of a predefined security document type. The raw data comprises grid-arranged measurement locations on the security document and preferably have the same resolution for a predefined area. Preferably, the measurement locations correspond to the same sections of the respective security document; more preferably, the entire security document; and most preferably, only the entire security document. The measurement locations can be identified by their position on the grid, by corresponding position data, or by their type and / or sequence in memory. The sensor values ​​can include intensity values ​​representing ultrasonic transmission.

[0065] In one embodiment, the evaluation procedure can include classifying the asset to determine its state or deviation from a permissible state, with the classification being directly dependent on the parameters to be determined. The parameters of the evaluation procedure can be optimized directly based on the distortion-reduced training data. The parameters of the evaluation procedure can include weights for linking components of the evaluation procedure and / or hyperparameters (Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE). The choice of an optimization function for the weights of the evaluation procedure depends on a specific classification model of the evaluation procedure. For example, a support vector machine (SVM) or a convolutional neural network can be considered as a classification model.The hyperparameters of the classification model can be determined using grid search and / or random search in a cross-validation procedure.

[0066] In the case of a single-function vector (SVM), the parameters can represent weights for linking the support vectors. In the case of a neural network, especially a convolutional neural network (CNN), the parameters can represent weights for linking nodes of successive network levels.

[0067] Determining the parameters can include at least one data preparation step. In particular, a predetermined number of classification features can be determined or extracted based on the sensor values. Generating the classification features can involve a dimensionality-reducing analysis, especially principal component analysis (PCA) or linear discriminant analysis (LDA). Principal component analysis, in particular, allows features to be extracted from the grids, enabling particularly efficient and reliable verification of the value documents.

[0068] A principal component analysis can include determining the mean and covariance matrix of the generated training data. Furthermore, the principal component analysis can include determining at least a predetermined number of eigenvectors of the covariance matrix. Based on this predetermined number of eigenvectors, a transformation matrix can be created. The principal component analysis can also include determining approximate sensor values ​​by applying the transformation matrix to the sensor values ​​of the raw data and determining a residual between the acquired sensor values ​​and the approximate sensor values.

[0069] In a further aspect, a procedure for evaluating securities of a specified type, in particular banknotes, is provided. The procedure involves the use of an evaluation method, whereby the parameters of the verification model are determined as described above. Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE

[0070] The method may further include the acquisition of a large number of spatially resolved raw data for a security document to be examined, using at least one sensor device. The raw data may include sensor values, each assigned to a location on the security document. The method may further include the extraction of at least one (classification) characteristic for the security document based on the sensor values. The classification characteristics may be generated as described above. The method may also include the evaluation, in particular the classification, of the security document using the at least one characteristic to determine a condition of the security document or a deviation from a permissible condition.

[0071] In a further aspect, a method for processing, in particular checking and / or counting and / or sorting and / or destroying, valuables of a specified type, especially banknotes, is provided. The method comprises transporting the valuables individually or in batches past at least one sensor device. The method further comprises capturing, during transport, spatially resolved raw data of the valuables by the at least one sensor device. The raw data comprises sensor values, each assigned to a specific location on the valuables. The method further comprises processing the raw data, preferably in real time, using a procedure described above. Depending on the result of the check, the valuables can be sorted and / or destroyed.

[0072] In another aspect, a device for evaluating securities of a predefined type, particularly banknotes, is provided using distortion-reduced training data and / or a test model. The training data was generated beforehand using a method described above. The parameters of the evaluation method were also determined using a method described above.

[0073] The device comprises a storage unit in which instructions for executing an evaluation procedure are stored. The device further comprises an evaluation unit configured to execute, by means of the instructions, a procedure for processing, in particular checking and / or counting and / or sorting and / or destroying, valuable documents as described above, preferably in real time. The device also comprises at least one sensor unit connected to the evaluation unit via a Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE

[0074] The data interface is coupled. The at least one sensor device is configured to capture raw data from the valuable documents, in particular by changing a measurement position on each of the valuable documents along one of at least two tracks, preferably parallel tracks. The at least one sensor device is further configured to transmit the raw data to the evaluation device, preferably in real time.

[0075] The evaluation unit can comprise a processor that is connected via data connections to the storage unit and to at least one sensor unit. Within the scope of the present invention, a processor is understood to mean not only a single processor with one or more cores, but also a system of coupled processors. The storage unit, the data connections, and the processor can form parts of a data processing unit of the evaluation unit.

[0076] In a further aspect, a device for processing, in particular checking and / or counting and / or sorting and / or destroying, valuables of a specified type, especially banknotes, is provided. The device comprises a feeder for feeding individual or single valuables to be processed. The device further comprises an output device with at least one output section for receiving processed valuables. The method also includes a transport device for transporting individual or single valuables from the feeder to the output device. Furthermore, the device comprises an evaluation device as described above. The at least one sensor device is arranged along a transport path for the valuables to be checked.

[0077] In another aspect, for each of the above-described procedures, a computer program with program code means is provided to carry out the respective procedure when the respective computer program is executed on a computer.

[0078] In a further aspect, for each of the above-described methods, a computer-readable data carrier containing program code is provided, which can be executed by a computer, so that the computer carries out the respective method. Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE

[0079] The invention, or further embodiments and advantages thereof, are explained in more detail below with reference to drawings, which only depict embodiments of the invention. Identical components are identified by the same reference numerals in the drawings. Elements drawn with dashed lines are considered optional.

[0080] The drawings are not to scale, and individual elements may be exaggerated or simplified. They show:

[0081] Fig. 1 shows an example of a schematic setup of a document processing device;

[0082] Fig. 2 shows an example of a roughly schematic representation of a device for generating distortion-reduced training data for determining parameters of a test model for checking securities;

[0083] Fig. 3a shows an example of a grid with sensor values ​​before a self-calibration of the at least one sensor device;

[0084] Fig. 3b shows an example of a grid with sensor values ​​after self-calibration of the at least one sensor device;

[0085] Fig. 4 shows a schematic representation of a method for generating distortion-reduced training data;

[0086] Fig. 5 is an example of a grid with sensor values ​​after changing the sensor values ​​according to one aspect of the invention;

[0087] Fig. 6 shows four orientations of an exemplary security document;

[0088] Fig. 7 shows an example of a flowchart of a procedure for determining parameters for a procedure for evaluating securities; and

[0089] Fig. 8 is an example of a flowchart of a procedure for verifying securities using an evaluation method, Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE

[0090] Parameters were determined using distortion-reduced training data.

[0091] A document processing device 10 in Fig. 1, in the example a device for processing documents 12 of a given document type in the form of banknotes, is designed for sorting documents 12 depending on the condition determined by means of the document processing device 10 and the authenticity of processed documents checked by means of the document processing device 10.

[0092] It has a feed unit 14 for feeding valuable documents 12, an output unit 16 for delivering or receiving processed, i.e., sorted, valuable documents, and a transport unit 18 for transporting individual valuable documents from the feed unit 14 to the output unit 16. In this example, the feed unit 14 includes an input tray 20 for a stack of valuable documents and a separator 22 for separating valuable documents 12 from the stack of valuable documents into the input tray 20 and feeding the individual valuable documents to the transport unit 18. In this example, the output unit 16 includes three output sections 24, 25, and 26 into which processed valuable documents can be sorted according to the intermediate result of the processing, in this example, verification. In this example, each of the sections includes a stacking tray and a stacking wheel (not shown) by means of which fed valuable documents can be placed in the stacking tray.In other embodiments, a dispensing section can be replaced by a device for destroying banknotes. The transport device 18 has at least two, in this example three, branches 28, 29 and 30, at the ends of which one of the dispensing sections 24, 25 and 26 respectively is arranged, and at the junctions via switches 32 and 34 controllable by signals, by means of which valuable documents can be fed to branches 28 to 30 and thus to dispensing sections 24 to 26 depending on the signals.

[0093] A sensor device 38 is arranged on a transport path 36 defined by the transport device 18 between the feed device 14, more precisely the singulator 22 in this example, and the first switch 32 after the singulator 22 in the transport direction T. This sensor device detects the physical properties of the valuable documents as they pass by and generates sensor signals representing the detection results, which comprise sensor data or raw data and represent sensor values. In this example, the sensor device 38 has at least two sensor elements 42, as well as other sensors 40, symbolized only by an additional box, for the physical properties of a valuable document. A control and evaluation device 46 is connected to the sensor device 38 and the transport device 18, in particular the switches 32 and 34, via signal connections.In conjunction with sensor device 38, it classifies a valuable document into one of predefined sorting classes based on the signals or sensor data from sensor device 38. These sorting classes can be predefined, for example, based on a status value determined by the sensor data and a authentication value also determined by the sensor data. Status values ​​can be, for example, "circulateable" or "not circulateable," and authentication values ​​can be "forged," "suspected of forgery," or "genuine." Valuable documents with adhesive strips, for example, are "not circulateable." In some cases, an adhesive strip can also lead to suspicion of forgery. Depending on the determined sorting class, it controls the transport device 18, specifically the switches 32, by issuing control signals.34 such that the document is output to an output section of the output device 16 assigned to that class, according to its sorting class determined during classification. The assignment to one of the predefined sorting classes, or the classification itself, is based on criteria specified for assessing the condition and authenticity, which depend on at least some of the sensor data.

[0094] The sensor elements 42 can be converters or ultrasonic transducers of an ultrasonic sensor.

[0095] The control and evaluation unit 46 has, in particular, at least one corresponding interface 44 for the sensor device 38 or its sensors, especially the sensor elements 42, a processor 48 and a memory 50 connected to the processor 48, in which at least one computer program with program code is stored. During the execution of this program, the processor 48 controls the device and evaluates the sensor signals of the sensor device 38, in particular to determine a sorting class of a processed security document. Furthermore, program code is stored in the memory. During the execution of this program code, the processor 48 controls the device and, according to the evaluation, controls the transport device 18. Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE

[0096] The interface 44, the processor 50, and the memory 48, or a section of the memory 48 in which a corresponding computer program and process parameters are stored, constitute an evaluation unit 47 within the meaning of this disclosure. In this example, the evaluation unit 47 evaluates the signals from the sensor unit 38. The processor 50 and other sections of the memory 48 can also perform other functions, in this example, the control of the document processing device 10.

[0097] The sensor elements 42 are designed to detect the ultrasonic transmission of a valuable document as it is transported past the sensor unit 38 by means of the transport device 18. The evaluation unit 47 is configured to further process and evaluate the detected sensor data.

[0098] Depending on the document properties, the control and evaluation unit 46, more precisely the evaluation unit 47, uses sensor data from the various sensors to determine in partial evaluations whether the determined document properties provide an indication of the document's condition or authenticity. Subsequently, corresponding data can be stored in the control and evaluation unit 46, for example in memory 50, for later use. Based on the partial evaluations, the control and evaluation unit 46 then determines a sorting class as the overall result of the inspection according to a predefined overall criterion and, depending on the determined sorting class, generates the sorting or control signal for the transport unit 18.For processing valuable documents 12, valuable documents 12, either stacked or individually inserted into the input tray 20, are separated by the singulator 22 and fed individually to the transport unit 18, which transports the separated valuable documents 12 past the sensor unit 38. This unit records the properties of the valuable documents 12, generating sensor signals that reflect the properties of the respective valuable document. The control and evaluation unit 46 records the sensor signals or data, determines a sorting class for the respective valuable document based on these signals (in this example, a combination of a security class and a condition class), and, depending on the result, controls the switches so that the valuable documents are transported to an output section assigned to the respective sorting class. TI.

[0099] Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE

[0100] The evaluation unit 47, together with the sensor unit 38, forms an example of an evaluation device for evaluating securities of a predefined type based on an evaluation procedure. Accordingly, the computer program contains instructions for executing a procedure for evaluating securities of a predefined type, in particular banknotes, using the specified parameters. In the evaluation procedure, raw data is acquired by means of the sensor unit 38 on a grid of a security to be examined and made available in the evaluation unit 47 in a corresponding section of the memory 50. The raw data is then evaluated using an evaluation procedure that includes parameters determined using training data, in particular data generated by a procedure described below.

[0101] Fig. 2 shows an example of a roughly schematic representation of a device for generating distortion-reduced training data for determining parameters of a method for evaluating securities.

[0102] The device 70 can be a data processing device with a storage device 72 for storing raw data of the specified value document type as training data. The device 70 is configured to execute a method for generating distortion-reduced training data, as described below, and to store the generated training data in the storage device 72. For this purpose, the device can have at least one processor 74, which is connected to the storage device 72 via a data connection, and a program memory 76, also connected to the processor 74 via a data connection, in which program code is stored. When this program code is executed, the device uses the processor 74 to execute the method for generating the distortion-reduced training data described below. In other embodiments, the program memory 76 can also be formed by a section of the storage device 72.The adaptation device 70 may also have a data interface (not shown in the figure), for example a network card, via which generated training data stored in the storage device 72 and / or parameters of the evaluation procedure determined therefrom can be transferred to another device. Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE.

[0103] Fig. 3a shows an example of a grid 60 before self-calibration of the at least one sensor device 38. The sensor device 38 can, in particular, comprise an ultrasonic sensor that recalibrates itself at regular intervals during operation. The grid 60 is not a physical grid of a value document 12 but a digital, virtual grid 60 that serves to address the measurement locations on the value document 12 or the training value document.

[0104] Each calibration of the sensor device 38 can put the sensor device 38 into a first state or a second state. The first state and the second state can be understood as post-calibration states. In the first state, the sensor device 38 can acquire sensor data from valuable documents, which will yield results of the evaluation process with an expectedly high quality. Sensor data acquired in the second state leads to results of the evaluation process with unexpectedly lower quality. However, the states after a calibration of the sensor device 38 are neither controllable nor predictable, so only a drop in the quality of the evaluation results can subsequently indicate the second state of the sensor device 38. By changing the sensor values, certain parameters can abstract the evaluation process from the different post-calibration states of the sensor device 38, i.e.,, the quality of the evaluation results remains predictably high, regardless of any interim recalibrations of the sensor device 38.

[0105] The grid 60 sketched in Fig. 3a may, for example, have been recorded after an initial calibration of the sensor device 38. The grid 60 may be divided into a predetermined number of cells, which are represented as square boxes in Fig. 3A. Alternatively, the cells may have a rectangular or polygonal shape. The grid 60 in Fig. 3A has seven parallel tracks of cells, with a number of cells between the endpoints of each track omitted for clarity. A track 61 is shown in Fig. 3A as an example by a dashed frame. The omitted cells are symbolized by three dots per track. The grid 60 extends along two mutually perpendicular, intersecting directions x and y. In the example in Fig. 3A, the x-direction is chosen to follow the transport direction T of the security document. Each track may have the same number of cells as the other tracks.Each cell can be assigned a sensor value. The sensor values ​​can represent an ultrasonic transmission of the security document 12 at the measurement location defined by the coordinate of the respective pixel (Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE). The resolution of the grid 60 perpendicular to the transport direction, i.e., the spacing of the tracks, is determined by the number and arrangement of the sensor elements, preferably ultrasonic transducers. The resolution of the grid 60 in the transport direction is determined, among other things, by the sampling rate and the transport speed of the security documents.

[0106] Fig. 3b shows an example of a grid 60 after self-calibration of the at least one sensor device 38.

[0107] The grid 60 in Fig. 3B was recorded from the same value document 12 as that in Fig. 3A and preprocessed using the same steps. Nevertheless, in the example in Fig. 3B, sensor values ​​were recorded for two of the tracks that differ considerably from the sensor values ​​of the same tracks in Fig. 3A. These tracks are shown in Fig. 3B as diagonally hatched to the right and left, respectively. The only difference between the recordings in Fig. 3A and Fig. 3B is a self-calibration of the sensor elements 42 of the at least one sensor device 38. Therefore, the difference in the sensor values ​​of the two affected tracks between Fig. 3A and Fig. 3B is attributed to the self-calibration of the sensor elements 42, which is necessary for operational reasons. However, not every self-calibration of the sensor elements 42 produces this difference in the sensor values. The difference is neither controllable nor predictable.After recalibration, the observed differences in sensor readings for at least one of the two tracks may disappear, at least partially, and / or new differences may appear in other tracks. These differences in sensor readings can occur randomly. However, these differences are not negligible and can lead to fluctuating results from a predetermined evaluation procedure, for example, causing a sudden deterioration in the quality of the evaluation result.

[0108] Differences in sensor readings can occur lane by lane, i.e., in each cell of a lane. Therefore, differences in sensor readings can also occur in the lane averages. In other words, the differences in sensor readings that may occur after self-calibration of the respective sensor element can be stable or constant over time. Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE

[0109] Since the self-calibration of the sensor device with its implications for the sensor values ​​during operation is required, a data set of grids 60 of several value documents 12 is considered distorted if it does not take into account the effect of random differences in the sensor values ​​after a self-calibration.

[0110] Fig. 4 shows a schematic representation of a method for generating distortion-reduced training data according to one aspect of the invention.

[0111] In a first step, Sl, raw data from training value documents are received. This raw data may have been acquired and stored, particularly after the initial calibration of the sensor device 38. It can be ensured that the raw data was acquired in an initial state of the sensor device, i.e., with perfect calibration. The raw data is spatially resolved, as shown, for example, in Fig. 3A. The raw data comprises sensor values ​​that represent a physical property of the (training) value document. For this purpose, it is advantageous that the training value documents consist exclusively of value documents without extraneous objects. The physical property could, for example, be the ultrasonic transmission of the value document, representing its thickness.For this purpose, the sensor device 38 can comprise an ultrasonic sensor that detects the training value document as it passes by in the transmission direction at a predetermined measurement frequency. The sensor device 38 can comprise at least two, preferably 4-16, more preferably 8-12 sensor elements, which can be controlled substantially synchronously. This allows the sensor elements 42 to detect a set of measurement locations on a grid 60 of a training value document based on a single pass of the training value document. The sensor elements 42 can detect sensor values ​​along parallel tracks of the training value document.

[0112] In a second step, S2, a compensation value C is provided based on a compensation distribution. The compensation value C can be a random number distributed according to the compensation distribution. The compensation distribution is preferably a mean-free normal distribution with a standard deviation in the range of 10 to 100 measurement units of the sensor value, more preferably in the range of 30 to 70 measurement units, and even more preferably in the range of 50 measurement units. In one embodiment, a compensation value C can be determined for each track; that is, each track in the grid 60 can be assigned its own compensation value C. The second step S2 can be executed once for the entire training dataset or for each training value document.

[0113] In a third step, S3, the acquired sensor values ​​are modified based on the determined compensation value C. In one embodiment, the sensor values ​​can be modified per track, i.e., using the respective compensation value C. By selectively modifying the sensor values, the effect of a random difference in the sensor values, for example due to self-calibration of the sensor device 38, can be simulated, so that the training data generated using this method exhibit less data distortion than conventionally generated training data without modifying the sensor values.

[0114] In one embodiment, the second step S2 and the third step S3 can be repeated. This means that the received sensor values ​​can be modified based on a further determined compensation value C' (not shown). In this way, the training data can be doubled, quadrupled, increased tenfold, or multiplied further, i.e., artificially enriched.

[0115] In an optional fourth step, S4, the generated training data can be provided to a method for determining the parameters of an evaluation procedure. The generated training data can be stored temporarily or permanently. Temporary storage of the generated training data, particularly in volatile memory, allows the method to be integrated into the adaptation procedure for parameter determination. Permanent storage enables access to the distortion-reduced training data from other devices at any later time.

[0116] Fig. 5 shows an example of a 60' grid after changing the sensor values ​​according to one aspect of the invention.

[0117] In grid 60' of Fig. 5, the sensor values ​​are modified by adding a track-specific compensation value CJ with J=1, ..., 7, i.e., Gl, C2, C3, C4, C5, C6, C7. Likewise, the compensation values ​​Gl, C2, C3, C4, C5, C6, C7 could be subtracted from the respective sensor value – the crucial point is that the compensation values ​​CI, C2, C3, C4, C5, C6, C7 are used in the same way. The compensation values ​​CI, C2, C3, C4, C5, C6, C7 can be determined and applied on the fly, i.e., held in non-persistent memory. Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE

[0118] The compensation values ​​Cl, C2, C3, C4, C5, C6, and C7 can be discarded after use. Looking back at the originally recorded grid 60 from grid 60' may be impossible. However, in one embodiment, both the originally recorded grid 60 and the respective grid 60' with modified sensor values ​​can be stored for each training data document. This allows the received or, if applicable, recorded training data to be freshly or separately modified during future determinations of the parameters of an evaluation procedure.

[0119] Fig. 6 shows four orientations of an exemplary security document 12. The front side 12a of the security document 12 can be captured in two orientations: in the reading direction and rotated 180 degrees. Similarly, the back side 12b of the security document 12 can be captured in two orientations: in the reading direction and rotated 180 degrees. Therefore, the first step, S1, of the procedure described with reference to Fig. 4, for each training security document, can comprise the acquisition of a multitude of spatially resolved raw data in each of the four orientations shown. In other words, the training data, starting from each training security document, can comprise 4 grids of 60° with compensated sensor values.

[0120] Fig. 7 shows an example of a flowchart of an adaptation procedure for determining parameters for an evaluation procedure or verification model for evaluating / verifying value documents.

[0121] In an optional preparatory step, S4, which corresponds to the optional fourth step S4 of the procedure from Fig. 4, distortion-reduced training data is provided.

[0122] In the first step, S10, predetermined features, such as classification features, are extracted from the training data. In one implementation, the classification features can be extracted using a dimensionality-reducing analysis, such as PCA. The classification features can be extracted for each grid 60, particularly in parallel execution. The first step, S10, is therefore a data preparation step. Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE

[0123] In a second step, Sil, the parameters of a selected evaluation method are determined based on the extracted classification features. For this purpose, it is advisable that the training data consist exclusively of value documents without any foreign objects. In this case, the label for all grid 60, or rather its classification features, is identical when determining the parameters. The label can be set to "0" and represent the class "no suspicion of manipulation". The label can be included in the received training data and either transferred to the bias-reduced training data or added to it.

[0124] In an alternative embodiment, the composition of the training value documents can include individual instances containing foreign objects, such as adhesive tape. In this case, in addition to the classification features, a label vector is necessary to determine the parameters, whereby the label vector assigns a label to each training value document indicating whether the corresponding training value document contains a foreign object (e.g., label value "1") or not (e.g., label value "0").

[0125] Determining the parameters in the second step, SH, is usually based on optimizing the parameters with respect to the input data, but depends on the selected evaluation method. Examples of such methods include a support vector machine (SVM), a convolutional neural network (CNN), or similar approaches.

[0126] Optionally, the specific parameters can be saved to make them accessible for transfer to other document processing machines.

[0127] In an optional third step, S12, the evaluation procedure, configured by the specified parameters, is provided for evaluating a value document 12.

[0128] In one embodiment, 10 instructions can be stored in a program memory of the security document processing device. When executed by the processor of the security document processing device 10, these instructions can cause it to evaluate security documents 12 using the evaluation procedure defined above. Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE

[0129] Fig. 8 shows an example of a flowchart of a procedure for evaluating value documents using an evaluation method whose parameters were determined using distortion-reduced training data.

[0130] In a first step, S20, spatially resolved raw data for a security document to be examined are acquired using at least one sensor device. The raw data comprise sensor values, each assigned to a measurement location on the security document 12. The acquisition of the raw data can be carried out essentially analogously to the first step S1 of the procedure described with reference to Fig. 4, but can be limited to a single security document. In contrast to the acquisition of raw data from training security documents, it is sufficient for security documents to be acquired in any orientation. In one embodiment, the security documents 12 are transported past the at least one sensor device, either individually or intermittently. The first step can be performed in real time.

[0131] In a second step, S21, at least one feature, in particular a classification feature, is extracted from the sensor values ​​of the spatially resolved raw data. The feature extraction can preferably be performed analogously to the feature extraction in the first step S10 of the adaptation procedure described with reference to Fig. 7. The second step can be executed in real time.

[0132] In a third step, S22, the sensor values ​​on grid 60 of the corresponding value document 12 are classified using the extracted features, the determined parameters, and the predetermined evaluation procedure to determine a state or a deviation from a permissible state of the value document. The third step can be executed in real time.

[0133] In an optional fourth step, S23, the security document 12 is sorted and / or destroyed based on the result of the third step, S22. This means that the execution of steps S21 and S22 for a security document 12 is preferably completed before the security document 12 reaches the first controllable switch 32 of the security document processing device. Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE

[0134] Further embodiments differ from the embodiments described above in that the control and evaluation unit 46 is divided into two parts: an evaluation unit corresponding to the evaluation unit 47 and a separate control unit that receives signals from the evaluation unit and uses them for control.

[0135] Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE

[0136] Reference symbol list:

[0137] 10 document processing devices

[0138] 12. Security document with front side 12a and back side 12b

[0139] 14 Feeding device

[0140] 16 Output device

[0141] 18 Transport equipment

[0142] 20 Input fields

[0143] 22 singulators

[0144] 24, 25, 26 Output section

[0145] 28, 29, 30 Branch of the transport facility 18

[0146] 32 first switch

[0147] 34 second switch

[0148] 36 Transport route

[0149] 38 Sensor device

[0150] 40 Sensor

[0151] 42 at least two sensor elements

[0152] 44 Interface

[0153] 46 Control unit

[0154] 47 Evaluation unit

[0155] 48 processor

[0156] 50 storage

[0157] 60 (virtual) grid of a security document 12

[0158] 60' grid with modified sensor values

[0159] 61 lane

[0160] 70 Device for generating distortion-reduced training data

[0161] 72 Storage device

[0162] 74 Evaluation unit

[0163] 76 program memory

[0164] T Transport direction

[0165] C, CJ compensation value, J=l,... ,n, e.g. n=7

Claims

Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE Patent claims 1. Method for generating distortion-reduced training data for determining parameters of an evaluation procedure for securities (12) of a given type of securities, in particular banknotes, the procedure comprising: - Receiving a variety of location-resolved raw data for each training value document of a number of training value documents, wherein the raw data comprise sensor values, each assigned to a location on the training value document; - Determining a compensation value (C) for each of the sensor values ​​of at least two disjoint sections of the raw data based on a predetermined compensation distribution; and - Changing the sensor values ​​of the raw data based on the respective compensation value (C) to generate the distortion-reduced training data, wherein the evaluation procedure is set up to evaluate value documents (12) based on spatially resolved raw data captured therefrom, wherein the parameters of the evaluation procedure are determined based on the distortion-reduced training data.

2. Method according to claim 1, wherein the raw data assign a sensor value to each detected location on the value document, wherein changing the sensor values ​​comprises adjusting the sensor value, in particular by adding or subtracting at least one compensation value.

3. Method according to any of the preceding claims, wherein the disjoint sections of the raw data comprise at least two tracks (61), in particular parallel tracks, on the training value documents, the method further comprising: - Determine, for each of the at least two tracks (61), a compensation value (C) for the sensor values ​​based on the predetermined compensation distribution; and - Modify, for each of the at least two tracks (61), the sensor values ​​of the raw data based on the respective compensation value (C) to generate the distortion-reduced training data. Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE 4. Method according to one of the preceding claims, wherein changing the sensor values ​​comprises adding or subtracting the respective compensation value (C) to or from the sensor values ​​in a track-related manner.

5. A method according to one of the preceding claims, further comprising: recording the plurality of spatially resolved raw data for each training value document of the number of training value documents by means of at least one sensor device (38), preferably by changing a measurement position on the training value document along one of at least two tracks (61), in particular parallel tracks.

6. Method according to any of the preceding claims, wherein the predetermined compensation distribution is a mean-free distribution.

7. Method according to one of the preceding claims, wherein the predetermined compensation distribution has a standard deviation on the order of a systematic deviation of the sensor values ​​after calibration of a sensor device (38) for capturing the spatially resolved raw data.

8. Method according to any of the preceding claims, wherein the predetermined compensation distribution comprises or is a normal distribution.

9. Method according to one of the preceding claims, wherein the compensation value (C) represents a random difference between sensor values ​​for the same value document, which were recorded before and after a self-calibration of the at least one sensor device (38).

10. Method according to any of the preceding claims, wherein the sensor value represents a thickness or a basis weight of the respective location on the security document.

11. Method according to one of the preceding claims, wherein the at least one sensor device (38) comprises an ultrasonic sensor, wherein the ultrasonic sensor has at least two sensor elements (42), in particular ultrasonic transducers, wherein each sensor element (42) of the ultrasonic sensor is assigned to its own track (61), and wherein the ultrasonic sensor preferably measures an ultrasonic transmission at a predetermined ultrasonic frequency. Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE 12. Method according to any one of the preceding claims 3-9, wherein the at least two tracks (61) extend parallel apart along a longitudinal direction of the security document (12).

13. Method according to any of the preceding claims, wherein the training data further comprises raw data for each training value document in each of its orientations.

14. Method according to any of the preceding claims, wherein the number of training value documents comprises a number in the range of 50 to 5000, preferably 100 to 2500, more preferably 200 to 1000.

15. Method for determining parameters of an evaluation method for securities (12) of a given security document type, in particular banknotes, wherein the method comprises determining the parameters of the evaluation method based on distortion-reduced training data generated by a method according to the preceding claims.

16. Method according to claim 15, wherein the evaluation method comprises classifying the value document to determine a condition or deviation from a permissible condition of the value document (12), wherein the classification depends directly on the parameters to be determined.

17. Method for evaluating securities (12) of a given type of security document, in particular banknotes, wherein the method comprises the use of an evaluation method according to claim 15 or 16.

18. The method of claim 17, further comprising: - Recording a large number of spatially resolved raw data for a security document (12) to be examined by means of at least one sensor device (38), wherein the raw data comprise sensor values ​​which are each assigned to a location on the security document (12); - Extracting at least one feature, in particular a classification feature, for the value document (12) from the sensor values; Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE - Evaluating, in particular classifying, the security document (12) using at least one characteristic to determine a condition or deviation from a permissible condition of the security document (12).

19. Method for processing, in particular checking and / or counting and / or sorting and / or destroying, securities (12) of a given type of securities, in particular banknotes, the method comprising: - Transporting the valuable documents (12) individually or intermittently past at least one sensor device (38); - Capture, during transport, by means of at least one sensor device (38), spatially resolved raw data of the value documents (12), wherein the raw data comprise sensor values, each of which is assigned to a location on the value document (12); and - Processing the raw data, preferably in real time, using a method according to one of claims 17 or 18.

20. Device for evaluating securities (12) of a predetermined security document type, in particular banknotes, using distortion-reduced training data generated in a method according to one of claims 1 to 14 and / or using an evaluation method according to one of claims 15 or 16, the device comprising: a storage device (50) in which instructions for executing an evaluation method are stored; an evaluation device (47) configured to execute a method according to one of claims 17 to 19 by means of the instructions, preferably in real time;at least one sensor device (38) which is coupled to the evaluation device (47) via a data interface (44) and is configured to capture raw data of the value documents (12), in particular by changing a measurement position on each of the value documents (12) along one of at least two tracks (61), preferably parallel tracks, and to transmit the raw data to the evaluation device (47), preferably in real time.

21. Device for processing, in particular checking and / or counting and / or sorting and / or destroying, valuable documents of a specified valuable document type, in particular banknotes, comprising the device: a feeding device (14) for feeding individual or isolated Giesecke+Devrient Currency Technology GmbH, Case: 514634 DE, a processing unit (12) comprising a supply unit (16) with at least one output section (24, 25, 26) for receiving processed documents (12), a transport unit (18) for transporting individual or separate documents (12) from the supply unit (14) to the output unit (16), and an evaluation device according to the preceding claim, wherein the at least one sensor unit (47) is arranged on a transport path (36) of the documents (12) to be checked.

22. Computer program comprising program code means to perform the method according to any one of claims 1 to 14 when the program is executed on a computer.

23. Computer-readable data carrier containing program code that can be executed by a computer such that the computer performs a method according to any one of claims 1 to 14.

Images