The invention relates to a method and a device for classifying security documents such as banknotes.

Many central banks are concerned with sorting secu ^¬ rity documents, in particular with determining whether banknotes are suitable for recirculation, or rather should be shredded and replaced by new ones. Obviously, more fre ^¬ quent recirculation reduces the printing costs and environmental burden. Given the huge amounts of different kinds of banknotes in circulation for even small coun- tries, determining the fitness of banknotes poses a seri ^¬ ous technical challenge in terms of processing speed and accuracy .

A device and a method for sorting banknotes is dis ^¬ closed in applicant's European patent EP-B1-1043700.

It is an object of the present invention to provide an improved device and method for classifying security documents such as banknotes. SUMMARY OF THE INVENTION

The object of the invention is met by providing a method for classifying security documents, such as bank- notes, comprising the steps of:

during a training phase:

al) obtaining at least two digital training images of at least two respective training security documents;

a2) dividing each digital training image into a set of predefined areas;

a3) determining at least one feature for each area of each digital training image;

a4) determining at least one most discriminating feature among all of the at least one feature and determining at least one threshold of the at least one most discriminat ^¬ ing feature;

during an operating phase:

bl) obtaining a digital image of a security document;

b2) determining at least one value of the at least one re- spective most discriminating feature;

b3) comparing the at least one value with the at least one respective threshold of the at least one most discriminat ^¬ ing feature; and,

b4) classifying the security document based on the com- parison.

According to the invention classification of the security documents (or assigning each security document to a class) may take place on the basis of the most discrimi ^¬ nating feature (s), while the most discriminating fea- ture(s) is/are determined on the basis of a set of train ^¬ ing security documents. Because of this, only a limited number (depending on how many most discriminating features are used in the operation phase) of features of the secu ^¬ rity documents that are to be classified, may need to es- tablished and this may increase the speed of the classifi ^¬ cation process. Since the features that are used to classify the se ^¬ curity documents are determined as the most discriminating features for the set of training security documents, the accuracy of the classifying process during the operating phase may be high. The accuracy may be further increased by increasing the number of most discriminating features, as is explained below.

Furthermore, because of the determination of the most discriminating features, high accuracy in the classifying process may be obtained with a limited number of training security documents.

In an embodiment of the method according to the in ^¬ vention, step a2) further comprises: providing all areas of the set of predefined areas with respective first area weighting factors; and in step a4) said at least one most discriminating feature and said at least one threshold is determined using said first area weighting factors.

It may be advantageous to weigh the different areas when determining the at least one most discriminating fea- ture and the at least one threshold. For example, one side of the security document or one specific area of that side may be more important or relevant with respect to classi ^¬ fying the security documents than others. For example, in classifying banknotes (an example of the security docu- ments) the area comprising the text with the value of the banknote may be more relevant than the area comprising the signature of the bank president.

In an embodiment of the method according to the in ^¬ vention, step a3) further comprises: providing all fea- tures of said at least one feature with respective feature weighting factors; and, in step a4) said at least one most discriminating feature and said at least one threshold is determined using said feature weighting factors.

Since some features may be more relevant or important with respect to classifying the security documents than others, it may also be advantageous to weigh the different features . In an embodiment of the method according to the in ^¬ vention, in step b4) the security document is classified based on the comparison using said first area weighting factors and/or said feature weighting factors.

It may be understood that weighing factors may be used for classifying the security documents, wherein the classification is based on the weighted combination of comparison of each of the at least one determined value with the at least one respective threshold.

It may be the case that the security document is classified as "fit" (or "unfit") when all of the at least one determined value are above their respective thresh ^¬ olds, when 50% (or another percentage) of all of the at least one determined value are above their respective thresholds, or when certain ones of the at least one de ^¬ termined value are above their respective thresholds.

In an embodiment of method according to the inven ^¬ tion, wherein step a4) further comprises:

- providing all areas of the set of predefined areas with respective second area weighting factors;

- determining a first most discriminating feature using said second area weighting factors;

- adjusting said respective second area weighting factors; and,

- determining a second most discriminating feature using the adjusted second area weighting factors.

For example, before determining the n-th (n being a natural number > 2) most discriminating feature, the sec ^¬ ond area weighting factors may be adjusted, such that the second area weighting factors of those areas that are cor ^¬ rectly classified by a combination of the n-1 most dis ^¬ criminating features may be decreased (or even be set equal to zero) and the second area weighting factors of those areas that are not correctly classified by a combi- nation of the n-1 most discriminating features may be increased . An advantage of adjusting the second weighting fac ^¬ tors is that areas that have been incorrectly classified so far are more significant or important when the next most discriminating feature is determined.

In an embodiment of the method according to the in ^¬ vention, the predefined set of areas comprises overlapping rectangular areas with various sizes and aspect ratios.

An advantage of this embodiment may be that it en ^¬ ables the determination of the same feature for different sized and overlapping areas. It may be the case that a feature is more discriminating when determined for an area Al of the training security documents than when it is de ^¬ termined for an area A2 of the training security docu ^¬ ments, wherein area A2 may contain area Al completely. In this way, an optimum area for which the feature is most discriminating may be determined.

In an embodiment of the method according to the in ^¬ vention, the at least one feature comprises at least one of:

- an average of an intensity I of said area; and,

- a standard deviation of the intensity I of said area.

The intensity I of an area of an image of a security document may be indicative of the state of the security document. For example, dirt on the security document may lower the intensity of an area of the image of the secu ^¬ rity document and/or the standard deviation of this inten ^¬ sity.

In an embodiment of the method according to the in ^¬ vention, the at least one feature comprises at least one of:

- an average of a red colour content R of said area;

- a standard deviation of the red colour content R of said area;

- an average of a blue colour content B of said area;

- a standard deviation of the blue colour content B of said area; - an average of a green colour content G of said area; and,

- a standard deviation of the green colour content G of said area;

Also the colour content of an area of an image of a security document may be indicative of the state of the security document, since certain kind of dirt may espe ^¬ cially lower the reflection of certain colours.

In an embodiment of the method according to the in- vention, the at least one feature comprises at least one of:

- an average of a yellow-blue colour content YB of said area;

- a standard deviation of the yellow-blue colour content YB of said area;

- an average of a red-green colour content RG of said ar ^¬ ea; and,

- a standard deviation of the red-green colour content RG of said area;

wherein YB is any linear combination of R, G and B, such as YB = R + G - 2B, and RG is any linear combination of R, G, B, such as RG = R - 2G + B.

Also linear combination of colour contents of an area of an image of a security document may be indicative of the state of the security document. For example, dirt (or a sebum deposit) may be mainly apparent in the blue chan ^¬ nel .

In an embodiment of the method according to the in ^¬ vention, the method comprises step b5) sorting the secu- rity document based on the classification.

In an embodiment of the method according to the in ^¬ vention, the at least one feature is normalized by an av ^¬ erage of the intensity I of a predetermined region of the digital training image. An advantage of this embodiment may be that variations in overall illumination during the scanning of the security documents may not influence the determination of a value of a feature. In an embodiment of the method according to the in ^¬ vention, the intensity I = R + G + B.

In an embodiment of the method according to the in ^¬ vention, the step al) comprises the steps of:

i) scanning the at least two respective training security documents to obtain at least two respective scanned im ^¬ ages;

ii) de-skewing and cutting said at least two respective scanned images to obtain at least two de-skewed scanned images; and,

iii) fitting each of the at least two de-skewed scanned images into a predefined rectangular shape to obtain the at least two digital training images.

In an embodiment of the method according to the in- vention, the step bl) comprises the steps of:

i) scanning the security document to obtain a scanned im ^¬ age ;

ii) de-skewing and cutting said scanned image to obtain a de-skewed scanned image; and,

iii) fitting each of the de-skewed scanned images into a predefined rectangular shape to obtain the digital image.

In a further embodiment of the method according to the invention, the step iii) of al) and/or step iii) of bl) comprises aligning a security document printed image within said predefined rectangular shape.

The predefined rectangular shape of step iii) of step al) may correspond to the predefined rectangular shape of step iii) of step bl) .

When security documents such as banknotes are

printed, a tolerance may be allowed in the exact position ^¬ ing of the printed images relative to the paper boundary. Therefore, it may advantageous to align a security docu ^¬ ment printed image within said predefined rectangular shape in stead of aligning the paper boundary of the secu- rity document within the rectangular shape.

In an embodiment of the method according to the in ^¬ vention, the at least one most discriminating feature com- prises at least 10 most discriminating features, or pref ^¬ erably 40 most discriminating features.

An advantage of using a higher number of most dis ^¬ criminating features may be an increased accuracy of the classification process. However, a higher number of most discriminating features may cause a longer processing time, during the training phase and/or during the operation phase. Furthermore, the increase of accuracy of a method using N+1 most discriminating features with respect to a method using N most discriminating features, de ^¬ creases with a higher N. Therefore, the number N of most discriminating features may be at its optimum at 40.

In an embodiment of the method according to the in ^¬ vention, the step b4) comprises classifying the security document as fit or as unfit.

In an embodiment of the method according to the in ^¬ vention, the step a2) further comprises determining the set of predefined areas on the basis of a contrast of the at least two digital training images.

When different kind of security documents are to be classified, the set of predefined areas may take into ac ^¬ count all different layouts of all possible security docu ^¬ ments. Using such a predefined set may cause the training phase to take a long time. It may therefore be advanta- geous to use a set of predefined areas that corresponds to the layout of the security documents to be tested.

On the basis of the contrast of the at least two digital training images information may acquired regarding the layout of the security documents to be tested. With this information, the predefined set of area may be deter ^¬ mined. For example, the predefined set may be selected from a set of areas that takes into account all different layouts of all possible security documents.

The object of the invention is also met by providing a device for classifying security documents, such as bank ^¬ notes, comprising: - a training scanner arranged for obtaining at least two digital training images of at least two respective train ^¬ ing security documents;

- a training image processing unit arranged for dividing each digital training image into a set of predefined areas and for determining at least one feature for each area of each digital training image;

- a training processing unit arranged for determining at least one most discriminating feature among all the at least one feature and at least one threshold of the at least one most discriminating feature;

- a scanner arranged for obtaining a digital image of a security document;

- an image processing unit arranged for determining at least one value of the at least one respective most dis ^¬ criminating feature;

- a classifying unit arranged for comparing the at least one value with the at least one respective threshold and for classifying the security document based on the com- parison.

In an embodiment of the method according to the in ^¬ vention, the training scanner is the scanner, and/or the training image processing unit is the image processing unit .

In an embodiment of the method according to the in ^¬ vention, the device further comprises a sorting unit, arranged for sorting the security document based on the classification.

The advantages of the embodiments of the device ac- cording to the invention may be similar or equal to the advantages of the embodiments of the method according to the invention, as is explained in this document.

The various aspects and features described and shown in the specification can be applied, individually, wher- ever possible. These individual aspects, in particular the aspects and features described in the attached dependent claims, can be made subject of divisional patent applica ^¬ tions .

BRIEF DESCRIPTION OF THE DRAWINGS

The invention will be elucidated on the basis of an exemplary embodiment shown in the attached drawings, in which :

Figure 1 schematically depicts the steps of the training phase of an embodiment of the method according to the invention;

Figure 2 schematically depicts the steps of the op ^¬ eration phase of an embodiment of the method according to the invention;

Figure 3 schematically shows an example of a scanned image ;

Figure 4 depicts an example of a digital (training) image ;

Figure 5 shows schematically an example of a division of the digital (training) image into a set of areas ac ^¬ cording to an embodiment of the invention; and,

Figure 6 schematically depicts a device for classify ^¬ ing security documents according to an embodiment of the invention .

DETAILED DESCRIPTION OF THE INVENTION

Classification of security documents such as banknotes may take place in order to determine whether they are fit or unfit for circulation in society. Other examples of security documents are licenses, such as a driver license, coupons and other document that are regularly ex ^¬ changed among users and may represent a certain economic value .

The soiling of a security document may be the main reason for classifying a security document as unfit. Other aspects of fitness of a security document may be stains and limpness, which show a high correlation with the level of soiling. In this document the classification of secu ^¬ rity documents may be particularly concerned with the classification of security documents with respect to their soiling .

For euro bank notes, it was concluded that the main soiling mechanism may be that fingerprint deposits cumu ^¬ lates and eventually forms a yellow/brownish layer of aged sebum. In addition, it may be the (gentle) touch of the human fingers causing soil (particularly sebum) adhesion on the elevated parts, the crumble- or fold lines, of the banknote, which may be revealing a structural yet in- homogeneous appearance.

Because of the in-homogeneous appearance of soil, it may be very difficult or impossible to describe the level of soiling of security documents by objective, quantized characteristics (or features) that are applicable for all kinds of security documents in order to classify them as fit or unfit.

In an embodiment of the invention, a number of most discriminating features are determined for a certain kind of security document during a training phase using a set of training security documents. This determination may take place by machine learning techniques. The most dis ^¬ criminating features are then used during an operation phase to classify security documents.

Classification may refer to assigning a security document to one or more of a set of classes. For example, the set of classes may consist of the class "fit" and the class "unfit" and in that case the classification refers to assigning a security document to the "fit" class or to the "unfit" class. But the set of classes may also com ^¬ prise three or more classes, for example "fit", "unfit, but repairable" and "unfit and unrepairable".

A characteristic of a (training) security document may be determined on the basis of an image of said secu ^¬ rity document. In this document, features of the image or parts of the image of the security document are determined rather than characteristics of the security document it ^¬ self. It is assumed that the features of (parts of) the image represent characteristics of the security document. And thus that the features may relate to the level of soiling of the security document.

First, embodiments of the training phase of the method according to the invention are described below. In the training phase a number of most discriminating features is determined. Figure 1 schematically depicts the steps of the training phase of an embodiment of the method according to the invention.

A first step 11 during the training phase may be ob ^¬ taining at least two digital training images of at least two respective training security documents. Each of train- ing security documents has been classified before the start of the training phase. For example, the training se ^¬ curity documents may have been manually classified by a group of experts.

The number of training security documents used in the training phase may be around 150 in each of the classes, for example 150 security documents that are classified as "fit" and 150 security documents that are classified as "unfit".

Step 11, obtaining the digital training images of the training security documents, may comprise the steps of:

- step 15, scanning the at least two respective training security documents to obtain at least two respec ^¬ tive scanned images;

- step 16, de-skewing and cutting said at least two respective scanned images to obtain at least two de-skewed scanned images; and,

- step 17, fitting each of the at least two de-skewed scanned images into a predefined rectangular shape to ob ^¬ tain the at least two digital training images.

In step 15 scanned images may be obtained for example by a (training) scanner. Figure 3 schematically shows an example of a scanned image 31. The image may be taken from the front or the back side of the training security docu ^¬ ment, however at least two images of the same side of the training security documents are required in the training phase .

In step 16 the scanned imaged may be cut along the lines of the security paper area 32. This security paper area may be obtained using an intensity threshold above a noise level. To reduce noise, the red, green and blue col ^¬ our content may be added together to form the intensity image I = R+G+B. In this way, signal-to-noise ratio may be optimized and this may yield most of the paper region.

Then the scanned image may need to be de-skewed and/or fitted, allowing determination of a feature over similar regions of the security document. As such, a box may be fitted around the security document, from which skew parameters can be estimated and a linear transforma ^¬ tion may map the pixels to a rectangular and fixed sized digital (training) image.

In an embodiment of the method according to the in- vention, the fitting of each of the de-skewed scanned im ^¬ age into a predefined rectangular shape to obtain the digital image comprises aligning a security document printed image within said predefined rectangular shape.

In general, a tolerance may be allowed in the exact positioning of the security document printed images (which may comprise offset and intaglio prints) relative to the paper boundary, when security documents are generated. To reduce the variation between regions introduced by the al ^¬ lowed tolerances in the printing process, it may be advan- tageous to align the security document printed images more accurate or precise. This may be achieved using the fol ^¬ lowing steps.

From a selection of training security documents that are classified fit, the one security document inducing the least amount of variation when being overlaid on the other security documents of the selection, may be determined. For this, a security document with minimum summed absolute colour difference between the pixels colour content of all other security documents may be taken, wherein the colour difference between two pixels may be considered to be the sum of the absolute differences between the three respec- tive colour contents. The resulting image may yield the typical (or modal) positioning of the security document offset layers within said selection. This resulting image may be used as a reference image for alignment of all other security documents.

Alignment of a given security document image may then proceeds as follows. The security document image is de- skewed as described above. After that, the image may be shifted in an x- and y-direction within an n x n

neighbourhood, and matched against the reference image for all possible shifts within the neighbourhood. The shift with minimum summed colour difference to the reference im ^¬ age may yield the best alignment between the given secu ^¬ rity document image and the reference image. In this way, security document images may be aligned to the major con- tent of the printing layers, rather than to the security document paper area.

In the above the step of obtaining at least two digi ^¬ tal training images of at least two respective training security documents during the training phase is described. It may be understood that similar steps and embodiments may also be applicable on the step of obtaining a digital image of a security document in the operation phase.

Figure 4 depicts an example of a digital (training) image 41, that may be used in an embodiment of the method according to the invention or by an embodiment of the de ^¬ vice according to the invention.

In the next step during the training phase, step 12 in figure 1, each digital training image is divided into a set of predefined areas, wherein the predefined set of ar- eas may comprise overlapping rectangular areas with various sizes and aspect ratios. A training image processing unit may be arranged for executing step 12. Figure 5 shows schematically an example of a division of the digital (training) image 41 into a set of areas 51 according to an embodiment of the invention. Although in figure 5 areas 51 have a rectangular shape, areas 51 may have a circular or any other two-dimensional shape. Areas 51 may or may not overlap each other. Areas 51 may have various sizes and may cover a large portion of the digital training image, for example an area 51 may cover the whole of the digital training image. The area 51 may cover a se- curity mark on the image, for example a depiction of the value that the document is representing or any other de ^¬ piction .

The set of predefined areas may comprise a set of random areas, i.e. a set of areas with random sizes and random positions. The set of predefined areas may comprise areas that have been selected for a certain kind of secu ^¬ rity document. The set of predefined areas may comprise areas that are assumed to be suitable for all kinds of se ^¬ curity documents.

In an embodiment of the invention, the set of prede ^¬ fined areas is determined on the basis of the digital training images during the training phase. On the basis of contrast or contrast patterns of the training images it may be determined which kind of security document is proc- essed and a set of predefined areas may be selected ac ^¬ cordingly.

A set of predefined areas may be generated, wherein the set comprises areas around regions with a high or a low contrast in comparison with other regions. Instead of contrast, also colour content patterns (of R, G and or B) of the training images may be used in determining or generating a set of predefined areas.

In the next step, step 13 of figure 1, at least one feature for each area of each digital training image is determined. A training image processing unit may be ar ^¬ ranged for executing step 13. Examples of features are: - an average of an intensity I of said area; - a standard deviation of the intensity I of said area, which may represent the contrast of the area;

- an average of a red colour content R of said area;

- a standard deviation of the red colour content R of said area;

- an average of a blue colour content B of said area;

- a standard deviation of the blue colour content B of said area;

- an average of a green colour content G of said area; - a standard deviation of the green colour content G of said area;

- an average of a linear combination of R, G and B, for example YB = R + G - 2B or RG = R - 2G + B . An advantage of (each of these two linear combinations may be that they may form together with the intensity the three orthogonal axis in a three dimensional colour space, and they may thus decorrelate the two chromatic information channels and the intensity channel.

It may be the case that digital (training) images of (training) security documents are obtained using electro ^¬ magnetic radiation with wavelength in the visible range or in the non-visible range. For example, using infrared (IR) or ultraviolet (UV) radiation. Therefore, in general, a feature may be an average or standard deviation of a col- our content, in which the colour is defined by a wave ^¬ length range and this wavelength range may be in the visi ^¬ ble spectrum, but may also be in the non-visible range, such as the IR or the UV range. And a feature may also be a (linear) combination of an average or a standard devia- tion of a colour content, in which the colour is defined by a wavelength range.

It may be the case that new security documents re ^¬ flect more light in a white region of the security docu ^¬ ment, for example in a region with a watermark. Therefore, the light intensity (and its standard deviation) of such an area of a digital image of a new security document may be higher with respect to an old security document. To counteract variations in overall illumination dur ^¬ ing the scanning of the security documents, for example due to accumulated dust on an image sensor of the scanner and variations in overall printing quality of the security document, any of the above listed features may normalized by an average intensity of a certain region or a colour content (for example R, G, and B, of that certain region) . This certain region may be its respective area, or any other area or may be the whole digital image.

The use of the blue colour content B (average or standard deviation) may by advantageous, since a sebum de ^¬ posit may be mainly apparent in the blue colour content.

In an embodiment of the invention, twelve features for each area 51 are determined during the training phase, being the average and the standard deviation of I, R, G, B, YB=R+G-2B and RG=R-2G+B. Furthermore, for each training security document, features from both the front and the back side of the training security document may be deter ^¬ mined. This may result in a large set of features, i.e. (the number of areas on the front and back side) x (number of features, for example 12) . As the examples of YB and RG show, a feature may also be a combination (for instance a linear combination) of the features described above.

However, only a small number out of the set of fea- tures may be used during the operation phase.

Instead of providing rules on how to classify secu ^¬ rity documents on the basis of one or more of these fea ^¬ tures, it may be advantageous to apply machine learning techniques to determine which features are most discrimi- nating.

In the next step, step 14 in figure 1, at least one most discriminating feature in the set of features (i.e. among all of the at least one feature) is determined and at least one threshold of these most discriminating fea- ture is determined. A training processing unit may be ar ^¬ ranged for executing step 14. Using the features of the training security documents and the known classification of the training security documents, it may be established which of the features is the most discriminating. For example, a feature (for in- stance: the average of the blue colour content of a cer ^¬ tain area 51, which is normalized by the average intensity of that area) may correctly classify 60% of the training security documents in a "fit" and an "unfit" class using a threshold of 0.4. The threshold may imply that a value be- low 0.4 corresponds to the fit class, while a value above 0.4 corresponds to the unfit class.

When no other feature in the set of features is bet ^¬ ter (i.e. correctly classifies a higher percentage of the training security documents), this feature may be identi- fied as the first most discriminating feature. The next best feature may then be identified as the second most discriminating features with its threshold and so on. In this way, a number of most discriminating features and their respective thresholds may be determined.

It may be the case that each area, into which each digital training image is divided, is provided with a re ^¬ spective first area weighting factor and/or a respective second area weighting factor.

The first area weighting factors may be used during the training and the operating phase. The first area weighting factors indicate the areas which are important for classifying the security documents, for example an area which depicts the denomination of a banknote.

The second area weighting factors may be used only during the training phase. The second area weighting fac ^¬ tors may be adjusted several times when the most discrimi ^¬ nating features are determined.

For example, after the first most discriminating fea ^¬ ture has been determined (or identified) , the respective second area weighting factors may be adjusted.

The second area weighting factors of those areas that are correctly classified by the first most discriminating feature may be decreased or may even be set equal to zero. The second area weighting factors of those areas that are not correctly classified by the first most discriminating feature may be increased.

The second most discriminating feature may then be determined with respect to all areas of all digital train ^¬ ing image, said areas having an adjusted second area weighting factor.

It may be understood that the above may also be ap- plied when determining (or identifying) the third, the fourth and following most discriminating features: before determining the n-th (n being a natural number > 2) most discriminating feature, the second area weighting factors may be adjusted, such that the second area weighting fac- tors of those areas that are correctly classified by a combination of the n-1 most discriminating features may be decreased (or even be set equal to zero) and the second area weighting factors of those areas that are not cor ^¬ rectly classified by a combination of the n-1 most dis- criminating features may be increased.

This may yield that, after selecting a first most discriminating feature, the areas of the training security documents will be reweighted such that already correctly classified areas become less important, whereas still in- correctly classified examples are emphasized in the selec ^¬ tion of the next most discriminating feature.

The combination of most discriminating features may be a linear combination of most discriminating features.

Figure 2 schematically depicts the steps of the op- eration phase of an embodiment of the method according to the invention.

In step 21 a digital image of a security document that is to be classified, is obtained. The step 21 may be executed similar to step 11. Likewise, step 21 may com- prise the above describes embodiments, for example the steps 15, 16 and 17. A scanner may be arranged for execut ^¬ ing step 21. In step 22 a value of the at least one respective most discriminating feature is determined. An image proc ^¬ essing unit may be arranged for executing step 22. Following the example described above, the value of the average of the blue colour content of the certain area 51, which is normalized by the average intensity of that area, may be determined for the security document that is to be classified. This value may be 0.6 in this example.

In step 23, the (determined) value of the at least one most discriminating feature is compared with the at least one respective threshold of the at least one most discriminating feature (in the example the threshold is 0.4) . And in step 24 the security document is classified on the basis of the comparison. In the example, the secu- rity document would be classified as unfit. A classifying unit may be arranged for executing steps 23 and 24.

In an embodiment of the invention, a step 25 is exe ^¬ cuted after step 24, wherein the security document is sorted based on the classification. In the example, the security document, which is classified as unfit, may be removed from circulation in the society. A sorting unit may be arranged for executing step 25.

For a most discriminating feature more than one threshold may be determined. For example, it may be the case that the security documents are to be classified in more than two classes, for example in three classes. In that case, a feature may have two thresholds. A value be ^¬ tween 0 and 0.3 may correspond to a "fit" class, a value between 0.3-0.4 may correspond to an "unfit, but repair- able" or "dubious" class and a value between 0.4-1.0 may correspond to an "unfit and unrepairable" class.

Figure 6 schematically depicts a device for classify ^¬ ing security documents according to an embodiment of the invention .

The device 61 for classifying security documents, such as banknotes may comprise: - a training scanner 62 arranged for obtaining at least two digital training images of at least two respective training security documents,

- a training image processing unit 63 arranged for divid- ing each digital training image into a set of predefined areas and for determining at least one feature for each area of each digital training image;

- a training processing unit 64 arranged for determining at least one most discriminating feature among all the at least one feature and at least one threshold of the at least one most discriminating feature;

- a scanner 65 arranged for obtaining a digital image of a security document;

- an image processing unit 66 arranged for determining at least one value of the at least one respective most dis ^¬ criminating feature;

- a classifying unit 67 arranged for comparing the at least one value with the at least one respective threshold of the at least one most discriminating feature and for classifying the security document based on the comparison.

In an embodiment of the device, the device further comprises a sorting unit 68, arranged for sorting the se ^¬ curity document based on the classification.

The training scanner 62 may be arranged for providing data regarding the at least two digital training images to the training image processing unit 63, which may be arranged for receiving said data. The training image proc ^¬ essing unit 63 may be arranged for providing data regard ^¬ ing the set of predefined areas and the at least one fea- ture for each area of each digital training image to the training processing unit 64. The training processing unit 64 may be arranged to receive said data.

The training processing unit 64 may be arranged for providing data regarding the at least one most discrimi- nating feature and the at least one threshold to the clas ^¬ sifying unit 67, which may be arranged to receive said data. The scanner 65 may be arranged for providing data regarding the at least two digital images to the image processing unit 66, which may be arranged for receiving said data.

The image processing unit 66 may be arranged for pro ^¬ viding data regarding the at least one value of the at least one most discriminating feature to the classifying unit 67, which may be arranged to receive said data. The classifying unit 67 may be arranged for providing data regarding the comparison to the sorting unit 68, which may be arranged to receive said data.

In an embodiment of the device, the training scanner 62 is the scanner 65; and/or the training image processing unit 63 is the image processing unit 66.

It is to be understood that the above description is included to illustrate the operation of the preferred em ^¬ bodiments and is not meant to limit the scope of the inven ^¬ tion. From the above discussion, many variations will be apparent to one skilled in the art that would yet be encom ^¬ passed by the spirit and scope of the present invention.