RU2700737C1 - Method of detecting foreign images on a banknote - Google Patents

Abstract

FIELD: monitoring devices.

SUBSTANCE: invention relates to a method of detecting foreign images on a banknote. According to the method, a digital image of the banknote is obtained and its compliance with one of the predetermined types and orientations is determined. Using the shape and location of the controlled area on the surface of the banknote, corresponding to the given type and orientation of the banknotes, an evaluation image is formed by binarisation of the monitored zone. Individual binarization threshold value for each pixel of the monitored area is calculated based on a binarisation parameter map predefined for a predetermined type and orientation of banknotes, and is associated with elements of said card with a known functional relationship. One element of the binarisation parameter map determines a binarization threshold value for a group of at least two neighboring pixels, and the known functional relationship is established taking into account the statistical distribution of pixel brightness in the digital image of the banknote. Then, an evaluation criterion is used to determine an alien image detection criterion based on an analysis of black pixels in said evaluation image, and based on said criterion, a decision is made on the presence of an extraneous image in the controlled area of the digital image of the banknote. Elements of the map of the binarisation parameter are selected so as to increase corresponding individual binarization threshold values and provide an acceptable low level of probability of a false decision on the presence of an extraneous image when performing said sequence of actions for banknotes of a predetermined type and orientation, which obviously do not contain extraneous images.

EFFECT: high sensitivity of the method of detecting foreign images with reduced parasitic sensitivity to geometrical distortions of scanning.

13 cl, 13 dwg

Description

The invention relates to the field of automated banknote verification. When monitoring banknotes in circulation, an important task is to find copies that have damage that make their further use undesirable. This task is called bankruptcy check and is performed on automatic counting and sorting machines.

During the decay test, the counting and sorting machine checks the banknote for a number of signs of decay: through holes, torn corners, torn edges, contrasting spots, patterns, inscriptions, scuffs of the paint layer and the like. To do this, when passing through the machine, the banknote is scanned from both sides, using various light sources. As a result, a plurality of digital banknote images are obtained corresponding to a particular side of the banknote and the radiation source. Further, banknote images are analyzed and as a result of the analysis, a decision is made on the presence or absence of certain signs of disrepair.

Signs of decay associated with damage to the base of the banknote are usually found as a result of the analysis of the image obtained in the transmitted radiation. This includes through holes, bent corners, and torn edges. To analyze surface damage, such as spots, patterns, inscriptions and abrasions of the paint layer, banknote images obtained in reflected light are usually analyzed. Since these damages can be on both sides of the banknote, in order to check their presence, it is necessary to analyze images that correspond to both sides.

Damage in the form of spots, inscriptions, drawings and stamp imprints can be attributed to the general type of extraneous images on the surface of the banknote. These extraneous images are superimposed on the normal image deposited on the surface of the banknote during its manufacture.

Therefore, when analyzing a digital image of a banknote, in order to find this type of damage, it is necessary to check the appearance of an extraneous image against the background of a normal image.

Known methods for searching for extraneous images based on pixel-by-pixel comparison of the digital image of the analyzed banknote with the digital image of the reference banknote. The difference between these two images, exceeding the predetermined limits, is considered as a sign of detection of an extraneous image. For example, in Japanese patent JP S5973758 (publ. 04/26/1984, IPC G07D 7/00) for the search for defects, the number of pixels of a digital image is calculated for which the difference with the reference image exceeds a predetermined difference threshold value. If the number of such pixels exceeds a predetermined limit value, then it is believed that there is an extraneous image on the banknote.

The above method has many disadvantages. If geometric distortions occurred while scanning the digital image under investigation, even in the absence of extraneous images, the light pixels of the image under study can shift and fall to the location of the dark pixels of the standard. Similarly, the dark pixels of the digital image under study due to the offset will be combined with the light pixels of the standard. As a result of comparing the digital image under study with the standard, for a large number of pixels, an excess of the allowable threshold difference value will be recorded, their number will exceed the allowable limit value, as a result of which a false decision will be made about the presence of an extraneous image. Since geometric distortions of scanning with varying degrees of probability occur in the mechanisms of almost any counting and sorting machine, pixel-by-pixel comparison leads to a noticeable percentage of false rejection of banknotes by the criterion of detecting an extraneous image. To reduce the rejection, it is necessary to make the mechanism for moving banknotes more complex and expensive in order to reduce geometric distortions during scanning.

Per-pixel comparison is a resource-intensive process. The machine has to store reference digital images of banknotes for all processed currencies and their denominations, as well as series of issuance, which differ in the design of the banknote. A significant amount of processor memory is used for this, which implements pixel-by-pixel comparison.

To correct the disadvantages of pixel-by-pixel comparison, methods are proposed in which, prior to the comparison, one way or another reduces the real spatial resolution of the image. The decrease in spatial resolution also reduces the effect of geometric distortion of the scan on the result of the method. For example, in the patent of the European Patent Office EP 2355056 (published on 08/10/2011, IPC G07D 7/00) the entire surface of the banknote is divided into pixel blocks of the same size, the average brightness level is calculated for each of them, and this value is assigned to the brightness level of a pixel in low-resolution digital image that replaces the corresponding unit. Next, the resulting low-resolution digital image is analyzed by comparing the brightness values of each pixel in it with an individual threshold value predefined based on the statistical processing of digital images of a plurality of normal banknotes. If the number of low-resolution pixels that have exceeded the threshold value exceeds the set limit value, it is concluded that there is an extraneous image on the banknote.

This method is similar to the previously described pixel-by-pixel comparison with the standard, with the difference that a low-resolution digital image is used for comparison. However, precisely due to a decrease in resolution and a corresponding increase in pixel size, the result of this method is much less dependent on the geometric distortion of the scan. From the point of view of processor memory consumption, an array of threshold values takes up significantly less space than a full digital image of a reference banknote, which can be considered another advantage of this method.

Unfortunately, this method, while reducing the frequency of false rejections of normal banknotes, also has a reduced sensitivity to extraneous images. This is due to the fact that extraneous images of small size or low contrast make too small a contribution to the brightness of the pixels of a digital image of low resolution, and this small contribution often does not lead to the brightness going beyond the set threshold value. In addition, the coarsening of the threshold value as a result of the statistical processing of many banknotes leads to the fact that on banknotes with the brightest pattern and paper, an extraneous image in total with the normal image also does not lead to the threshold being set. As a result, for many extraneous images, a sufficient number of low-resolution pixels are not accumulated whose brightness value is outside the acceptable threshold value, as a result of which the extraneous image is not detected by this method.

In part, the disadvantages of this method are corrected in US Pat. block. Due to this, even small-sized elements of the extraneous image have a significant impact on the pixel values of low resolution, which increases the sensitivity of the method. However, this increase in sensitivity occurs not only in relation to spots, inscriptions and drawings, but also in relation to scanning noise and minimal image defects. Such an indiscriminate increase in sensitivity leads to false rejections and adversely affects the applicability of this patent.

Another disadvantage of pixel-by-pixel comparison is related to differences between normal banknotes arising from the instability of the printing process and due to the allowable wear of the banknotes themselves. On normal banknotes, such characteristics as the optical density of the ink layer and the thickness of the thin lines of the pattern are subject to significant changes from copy to copy due to the instability of the printing process. When banknotes are printed from different copies of printing forms and on different equipment, the distance between the characteristic elements of the picture may differ markedly. With acceptable wear of banknotes, the paper darkens, and the optical density of the ink layer decreases. When all these factors are summarized in the banknote under study, when comparing its digital image with the standard, too many pixels can be found in which a difference is found above the permissible one. This also leads to false rejection according to the criterion for detecting an extraneous image. In patent EP 2355056 (published on 08/10/2011, IPC G07D 7/00) statistical processing allows you to set individual pixel threshold values in such a way as to avoid false rejection of normal banknotes due to the spread of their parameters: density of the paint layer, paper brightness, line thickness , distances between characteristic elements. At the same time, an increase in pixel threshold values leads to a decrease in the sensitivity of the method. At its core, the pixel threshold value determines the local sensitivity for a particular pixel, and acts in the same way as in the case of a general darkening of a banknote, or in the case of an extraneous image. Increasing the threshold value in order to compensate for the overall darkening of the banknote, at the same time, reduces the sensitivity to local darkening caused by an extraneous image.

In order to maintain sensitivity to extraneous images over a wide range of changes in the overall brightness level of the document, various techniques are used based on taking into account the general darkening of the paper and the general wear of the ink layer before searching for local differences from the reference digital image. For example, to compensate for the effects of differences in paper brightness and optical density of the ink layer between different copies of banknotes with a normal image, in the method according to US patent US 8805025 (publ. 03.10.2013, IPC G07D 7/182), a digital image of a banknote after scanning is subjected to normalization of gradation interval. As a result of normalization, the initial interval of pixel brightness values, depending on the brightness of the paper and the optical density of the ink layer, is linearly displayed on a fixed interval. Then, the digital image is binarized in accordance with such a threshold value of binarization, in which in the digital image as a result of binarization a predetermined percentage of black pixels is obtained in advance. Both of these actions, to a large extent, reduce the differences between the digital images of various copies of banknotes.

After binarization, the analyzed digital image is compared pixel by pixel with the reference binarized image. As a result of the comparison, a digital image is obtained in which black pixels appear only in those places where there was a white pixel in the reference digital image and black was in the analyzed one. The black pixels of the resulting digital image are essentially a binarized representation of the extraneous image. Further analysis of the resulting digital image reveals black areas that exceed the predetermined allowable size. When such an area is identified, it is concluded that there is an extraneous image.

This method was chosen as the closest prototype to the claimed invention. In the prototype, the problem of the influence of differences in the brightness of the paper and the density of the ink layer between the banknotes is solved. Storing a reference binarized digital image requires slightly less memory than storing a complete reference digital image.

However, in the prototype, in fact, the principle of pixel-by-pixel comparison is used, which entails a significant parasitic sensitivity to geometric distortions. In particular, with a small shift of the black area relative to each other in the digital image of a normal banknote and in the reference digital image, a parasitic black contour could appear on the resulting digital image, which is mistakenly perceived as an extraneous image. In order to avoid such an error, in the reference digital image, the black region is additionally somewhat expanded, which leads to the appearance of a dead zone of the method for extraneous images along the perimeter of the black region.

In addition, due to the establishment of a single threshold value of binarization for all pixels in a digital image, the prototype has a significant limitation of sensitivity to real extraneous images. In the darkened areas of the normal image of the banknote, the reference binarized digital image already has black pixels, therefore, in such areas it is impossible to detect the appearance of extraneous images, even if they are applied with the darkest paint. On the contrary, in the brightest margins of a banknote, an extraneous image made with dark gray paint may not reach the threshold for binarization and will also go unnoticed. This is due to the fact that the proportion of black pixels recommended by the authors when setting the threshold value is 10-30%, that is, the threshold value corresponds to the brightness level in the darkest part of the normal image. The prototype is good enough to detect only extraneous images painted with dark paint on areas of medium brightness.

Consider possible ways to correct the disadvantages of the prototype. To reduce the parasitic sensitivity to geometric distortions, one should completely abandon the pixel-by-pixel comparison with the standard. To ensure high sensitivity to extraneous images, it would be desirable to adapt the binarization threshold value to the features of the normal image in each small area of the banknote. This would allow to identify extraneous images on any background, on which they can only be seen.

The technical result of the claimed invention consists in increasing the sensitivity of the method for detecting extraneous images, while reducing its spurious sensitivity to geometric distortion of the scan. An additional result is to reduce the amount of memory required to store invariable data used to implement the method. These results are achieved due to the fact that in the method for detecting extraneous images on a banknote

receive a digital image of a banknote to be checked for extraneous images, and determine its compliance with one of the predefined types and orientations,

and further, using the shape and location of the controlled area on the banknote surface corresponding to this type and orientation of the banknotes, an estimated image is formed by binarization of the controlled area, with an individual binarization threshold value for each pixel of the controlled area

calculated on the basis of a binarization parameter map predefined for a predetermined type and orientation of banknotes,

and is associated with the elements of this map with a known functional dependence, moreover, one element of the binarization map determines the threshold value of binarization for a group of at least two adjacent pixels,

and the known functional dependence is established taking into account the statistical distribution of pixel luminances in the digital image of the banknote to be checked in order to reduce the effect of the brightness spread on the result of binarization on various banknote copies,

and then, an extraneous image detection criterion based on the analysis of black pixels in that evaluation image is applied to the evaluation image,

and based on this criterion, a decision is made on the presence of an extraneous image in the controlled area of the digital image of the banknote to be verified,

moreover, the elements of the binarization parameter map are selected in such a way as to increase the individual binarization thresholds corresponding to them, and at the same time, to provide an acceptable low level of probability of a false decision about the presence of an extraneous image when performing the above sequence of actions for banknotes of a predetermined type and orientation, obviously containing extraneous images.

The claimed method does not use pixel-by-pixel comparison. Instead, an individual binarization threshold value is calculated for each pixel. This threshold value is set by the binarization parameter map, in which each element sets the binarization threshold value for several neighboring pixels. Due to this, if the position of the graphic element of the banknote is shifted to neighboring pixels due to geometric distortion of the scan, the displaced pixel usually remains within the group, the binarization thresholds for which are set by the same element of the binarization parameter map. Accordingly, geometric distortions have relatively little effect on the result of binarization. This reduces the sensitivity of the method to geometric distortion of the scan.

The binarization parameter map contains fewer elements than there are pixels in the original digital image. This happens because one map element sets a threshold binarization value for several pixels of a digital image. As a result, the implementation of the claimed method requires a smaller amount of non-volatile memory than for known methods using pixel-by-pixel comparison. Indeed, with pixel-by-pixel comparison, it is necessary to non-volatilely store the reference digital image of the banknote, which contains more elements than the binarization parameter map.

In the inventive method, the binarization parameter map elements are selected in such a way as to increase the individual binarization thresholds corresponding to them, and at the same time, to provide an acceptable low level of probability of a false decision about the presence of an extraneous image for banknotes of a predetermined type and orientation that do not contain extraneous images. With increasing threshold value of binarization, the number of pixels increases, the value of which turned out to be lower than the threshold value.

Accordingly, the number of black pixels in the evaluation image also increases. An increase in individual binarization thresholds applied to the digital image of a banknote with a stranger image increases the sensitivity to the stranger image, since after such an increase the stranger image, even having a small optical density, appears in the evaluation image.

At the same time, each individual binarization threshold value does not increase excessively, so as not to cause black pixels to appear in the estimated image, which can cause false detection criteria for an extraneous image when checking banknotes called normal, which do not have an extraneous image. That is, the binarization parameter map elements are selected to simultaneously fulfill two requirements: increase individual binarization thresholds, and reduce to the acceptable level of probability of false positives of the extraneous image criterion on normal banknotes.

False triggering of the criterion in the absence of extraneous images cannot be completely excluded due to the variation in the parameters of banknotes of a certain type, due to the instability of the printing process and the allowable wear of banknotes in circulation. In addition, for the formation of a binarization parameter map, for purely practical reasons, you can use only a limited sample of the total number of banknotes in circulation. Therefore, the requirement for individual binarization thresholds is only about reducing the probability of a false decision about the presence of an extraneous image to a technically acceptable level, but not about completely eliminating this conclusion. The acceptability of the specified level of probability can be established in a practical way based on the requirements for cash turnover adopted in the circulation zone of a banknote of a certain type. In many countries, an acceptable level of false decisions about the presence of an extraneous image on a normal banknote is up to 5% of the total number of processed banknotes.

Note that the selection of elements of the map parameter binarization in the claimed method can be performed in various ways. Full optimization with the achievement of the best possible values of the map of the binarization parameter, in the strict mathematical sense, is a difficult task, both in terms of the development of the algorithm and the complexity of its computational implementation. From a technical point of view, it is reasonable to use simpler sequence of steps to select values of map elements, which, although they do not lead to complete and mathematically accurate optimization, but give a significant approximation to it. The result of the selection, which is obtained in this case, is usually called quasi-optimal, that is, close to optimal, which can be used as optimal for practical purposes.

In comparison with the prototype, where a constant threshold binarization threshold value is used for all image pixels, in the claimed method, threshold values are individually selected in accordance with the partitioning of the estimated image into groups. This allows you to select quasi-optimal threshold values for both dark and light sections of the banknote to provide them with increased sensitivity to extraneous images. The claimed method, in contrast to the prototype, provides the appearance of black pixels after binarization in place of the extraneous image, regardless of whether the extraneous image is in the dark or light region of the banknote pattern. Thus, an increase in sensitivity is achieved.

Given these considerations, in a preferred embodiment of the inventive method, in order to construct a map of the binarization parameter corresponding to a certain type and orientation of banknotes, preliminary processing of a set of genuine banknotes that do not contain extraneous images is performed in advance, in the following description such processing will be called training, during which receive a set of training digital images of a set of genuine banknotes that do not contain extraneous images for each predefined type pa and orientation,

and form a map of the binarization parameter on the basis of the obtained set of training digital images, for which a set of calibration areas is set that together cover the controlled area,

and also the target value of the proportion of black pixels, less than 40%, set such initial values of the map elements of the binarization parameter, which when binarizing each digital image from the training set, ensure that the proportion of black pixels in each calibration region exceeds the target value,

and further, when considering each calibration area on each of the training digital images in the set, by changing the binarization parameter map elements, the binarization threshold value for individual pixels is reduced, then and to the extent that it is necessary to reduce the proportion of black pixels within each calibration area to a level no higher than the target value.

In a preferred embodiment, to limit the possibility of false triggering of the extraneous image criterion, the target value of the proportion of black pixels is used. The outside image appears as additional black pixels in the evaluation image, which are subsequently discovered by the criterion. As the target value of the proportion of black pixels increases, the total number of black pixels in the evaluation image also increases. Too many black pixels obtained by binarizing a digital image of a banknote without an extraneous image causes a false positive. Therefore, for a given criterion, such a target value of the proportion of black pixels is selected that does not allow the probability of false positives to exceed an acceptable level. In practice, various criteria are applied that usually correspond to the requirements of the central bank in a certain type of banknote circulation zone. For a given criterion, the selection of the target value of the proportion of black pixels is most conveniently carried out experimentally using a representative sample of normal banknotes, gradually changing this target value of the fraction until the probability of false positives about the presence of an extraneous image is not higher than an acceptable level. The probability of a false positive should not be significantly lower than an acceptable level, since this leaves an unused opportunity to further increase the individual binarization thresholds and increase the sensitivity of the method.

For some criteria, it is possible to theoretically calculate the corresponding proportion of black pixels based on geometric considerations. However, the calculated value usually requires experimental verification and correction.

In a preferred embodiment, an increase in individual binarization thresholds is provided, and the probability of false positives of the extraneous image criterion on normal banknotes is reduced to an acceptable level. More specifically, the binarization thresholds for individual pixels are first increased to a deliberately unacceptably high level, and then reduced only to the extent necessary to reduce the proportion of black pixels within each calibration region. Due to this, the threshold value of binarization is maintained at the highest possible level at which, on normal banknotes, the target value of the proportion of black pixels in all calibration areas is not yet exceeded. Thus, on a normal banknote, the evaluation image contains a background density of black pixels in all areas of the banknote not exceeding the target value. This, in turn, provides an acceptable low probability of a false decision about the presence of an extraneous image on a normal banknote.

When there is a foreign image on the banknote being checked, its optical density additionally reduces the brightness of a part of the pixels in comparison with a normal banknote. After binarization, even a relatively small decrease in brightness leads to an increase in the number of black pixels in the corresponding region and the appearance of an extraneous image in it. That is, unlike the prototype, an extraneous image appears, even if it does not have a high optical density on the banknote.

Due to the limitation of the proportion of black pixels in the calibration region with the target value, a high sensitivity of detection of extraneous images is achieved. No matter how dark the picture of a normal banknote is in a certain area, after binarization, the proportion of black pixels in it will not exceed the target value. In this area of normal banknote, a significant proportion of white pixels remains. However, if there is an extraneous image in it, the pixels in the evaluation image corresponding to the extraneous image change their color to black and thus appear.

There are various quasi-optimal strategies for lowering the threshold binarization values to reduce the proportion of black pixels within each calibration region to a level no higher than the target value. In one implementation, when considering each calibration area, on each of the training images,

calculate the limit level of binarization for the region at which the target proportion of black pixels is achieved in this region,

and, to reduce the binarization threshold value for each pixel in this region, if the previously set binarization threshold value of a given pixel exceeds the limit binarization level, by changing the binarization parameter map elements, reduce it to the binarization limit level.

The described implementation is simple and allows you to reduce threshold values in any order of consideration of training images and areas on them. In other implementations of the claimed invention, more complex sequences of steps can be used to provide the necessary reduction in threshold binarization values. In general, the higher the binarization threshold value is set, the more completely the extraneous images of low optical density will appear in the estimated image. Therefore, in order to obtain the highest sensitivity of the method to such images, it is desirable to reduce the threshold values to the smallest possible degree, achieving as close as possible the background fraction of black pixels from below to the target value of the fraction. When implementing the inventive method, it is advisable to choose a sequence of steps that provides sensitivity to extraneous images of low optical density, acceptable in practice, with an acceptable level of computational complexity.

The extraneous image is transferred to the evaluation image in the form of additional black pixels, which, in fact, distinguish a normal banknote from a damaged one. Therefore, it is reasonable to call such additional pixels abnormal. To assess the number and location of anomalous pixels, an appropriate criterion for their calculation should be used. It is obvious to a specialist that if an extraneous image is transferred to an estimated binary image, then selecting a criterion for its detection is a typical engineering task. A description of the solution to this problem is available both in the prototype and in other solutions known from the prior art. For example, in the patent of the Russian Federation RU 2282895 (published on April 10, 2005, IPC G07D 7/12), clusters of anomalous pixels are searched by checking for other anomalous pixels among eight possible neighbors of each anomalous pixel. To evaluate clusters, the patent proposes several different criteria that allow us to conclude whether the cluster corresponds to one form or another of an extraneous image. Based on the numerical characteristics of the clusters, such as the number of clusters on the banknote, the number of anomalous pixels in the cluster and the ratio of the cluster perimeter to its area, a conclusion is made about the dilapidation of the banknote. The prototype uses a close approach, but the criterion of disrepair is the excess of a certain limit on the size of the area containing the extraneous image.

Most often, according to the requirements of the central bank, you need to check the presence on the banknote of an extraneous image of a given size, for example, a dark circle. In this case, it is necessary to sequentially go through the evaluation image with a test window of the same size, and, for each position of the window, count the number of black pixels that appear in it. If the number of black pixels that appear in the window exceeds a predetermined limit value corresponding to the largest allowable size of the extraneous image, the criterion is triggered and issues a conclusion about the presence of the extraneous image. When an extraneous image does not get into the window, the maximum possible number of black pixels in the window (the so-called background number of pixels) is determined by the target values of the proportion of black pixels, which are chosen significantly below the limit value. Other, more complex calculation criteria are possible, the choice of which depends on the statement of the task of checking for disrepair according to the specific requirements of the central bank.

When the banknote being verified turns out to be darker or brighter, or more contrasting or less contrasting than a typical banknote, this is reflected in the statistical distribution of pixel brightness in its digital image. In the claimed method, the binarization threshold value is formed taking into account the statistical distribution of pixel brightness of the digital image. This is done in order to reduce the effect of the spread of brightness on different copies of banknotes on the result of binarization. For a specialist, the possible ways of taking into account the statistical distribution of pixel brightness are obvious. For example, for a darker image of a banknote, which is expressed in a reduced mathematical expectation of the brightness distribution, it is necessary to further reduce the individual threshold values of all pixels relative to those calculated on the basis of the binarization parameter map for a typical banknote. Accordingly, for a lighter banknote having an increased mathematical expectation of the brightness distribution, it is necessary to increase the individual threshold values of all pixels. Accounting for the statistical distribution of brightness in the claimed method successfully solves the same problem as the normalization of the interval of brightness used in the prototype.

Thus, the claimed method retains the advantages of the prototype, but improves sensitivity when detecting extraneous images and reduces spurious sensitivity to geometric distortions.

The purpose of adjusting threshold values using calibration areas is to, for a normal banknote without extraneous images, obtain an approximately constant density of black pixels in all parts of the banknote. It is important that, in this case, large parasitic clusters of black pixels that are similar in shape and size to a foreign image do not appear in the valuation image of such a banknote. Spurious clusters occur at the places of the darkest elements of the banknote pattern, and it is impossible to completely avoid their formation. If the parasitic cluster reaches the size and shape at which the detection criterion gives a false decision about the presence of an extraneous image, then the claimed method does not work correctly and produces a false positive result.

As a result of the method, within each calibration region, such binarization thresholds are set that the proportion of black pixels in the region, for a normal banknote, does not exceed a predetermined target value, which is selected to be less than 40%. This limitation of the proportion of black pixels allows the formation of small spurious clusters, the size of which is much smaller than the size of the calibration region, but does not allow the formation of large clusters, comparable in size to the region itself. The largest area of a possible cluster within the calibration area is equal to the target value multiplied by the area of this area. Therefore, the size of the calibration region and the target value must be chosen small enough so that spurious clusters cannot exceed the maximum dimensions of the extraneous image set as a decay criterion.

The target value of the proportion of black pixels should be selected empirically in accordance with a given detection criterion. In general, due to this target value, it is necessary to provide an acceptable low level of probability of a false decision about the presence of an extraneous image for banknotes of a predetermined type and orientation that do not contain extraneous images. As experience shows, the optimal target value of the proportion of black pixels lies in the range from 10% to 30%, which, accordingly, limits the maximum cluster area in the range from 10% to 30% of the calibration area. When the target value of the fraction increases to 50% or more, small parasitic clusters touch and merge with each other, resulting in a single large cluster. In fact, there is an inversion of the structure of the estimated image when scattered white clusters remain on a black background. This causes false detection criteria and makes the method inoperative. The practical limit target value for the proportion of black pixels determined during the experiments is 40%.

There is another way for large parasitic clusters to appear, which manifests itself even at small target values of the proportion of black pixels. We are talking about such a mutual arrangement of the calibration areas and the banknote pattern, in which parts of one cluster are spaced apart from neighboring non-overlapping areas. Each part of the cluster is small enough and does not, in its calibration region, exceed the fractions of black pixels of the set target value. At the same time, the entire cluster can be so large that its size can trigger the decay criterion. In order to limit the possibility of a parasitic cluster occurring at the boundary, the calibration regions should be chosen with some overlap. The overlap, in this case, will relate to at least two calibration areas, for each of which, separately, the threshold values of binarization are adjusted according to the limiting level of binarization. All mutually overlapping regions are involved in choosing the binarization threshold value for the overlap, which reduces the likelihood of a large parasitic cluster forming at the junction of the regions.

Of greatest difficulty are parasitic clusters located at the junction of the corners of the regions, since their parts are distributed over all four adjacent regions and, thus, a fourfold increase in the cluster area is theoretically possible. For example, a cluster evenly distributed over four regions, with a target value of the proportion of black pixels of 25%, may have a total area equal to the area of the calibration region. To suppress the formation of such clusters, a set of calibration regions in which

calibration areas are rectangular with constant width and constant height,

and the set of these areas consists of two subsets,

so that these calibration areas in each subset are placed close to each other in the order of rows and columns,

moreover, the calibration regions of one subset overlap with the intersecting calibration regions of the other subset half the height and half the width of the region.

The described arrangement of the calibration regions ensures that the parasitic cluster located at the junction of the corners of the regions of one subset will, at the same time, be located in the middle of the region from another subset. The threshold binarization values established for the location of such a cluster will be determined by the limiting level of binarization of the region in the middle of which the cluster is located. Thus, when setting threshold binarization values in this region, the cluster size does not exceed the target value.

The problem of large spurious clusters located at the boundary of the calibration regions can be almost completely solved by creating a dense regular grid of overlapping rectangular regions. For example, you can place areas in the form of rows and columns with certain pitch values both in the direction of the row and in the direction perpendicular to it. When the step is chosen small enough in comparison with the size of the region, then for almost any cluster location there is such a gauge region that the cluster is near its center. Adjusting the binarization thresholds for this region will not allow the cluster to go beyond the size specified by the target value of the proportion of black pixels. The disadvantage of this solution can be considered a significant increase in the number of calibration areas and the computational cost of training. By changing the step size, a technically justified compromise can be reached between the training time and the degree of suppression of large stray clusters.

It is also possible to implement a method in which the overlapping of the calibration regions is not used, and in order to combat the formation of large clusters, the size of the rectangular calibration region is chosen to be sufficiently small. We give an example of choosing the size of a rectangular region when the maximum allowable size of a spurious cluster is given by its area. To select the size of the calibration region, one should consider the already mentioned least acceptable case, when a large cluster is formed at the junction of the corners of the four calibration regions, so that in each region there is one quarter of the cluster area. The area of the part of the cluster within the same area does not exceed the area of the area multiplied by the target value of the proportion of black pixels. Therefore, the area of the calibration region must be selected less than one quarter of the maximum allowable cluster area divided by the target value of the proportion of black pixels.

One of the possible implementations of the invention involves taking into account the statistical distribution due to the fact that

for the banknote to be checked, a characteristic level is calculated such that the fraction of pixels with equal or lower values in the digital image of this banknote is the target value of the proportion of black pixels predefined for this type of banknote,

and the value supplied by the functional dependence is proportional to the characteristic level.

In mathematical terms, the characteristic level is the quantile of the histogram of the digital image of the banknote, and the limit of binarization is the quantile of the histogram of the calibration region of the digital image, and these quantiles are calculated for the same predefined target value of the proportion of black pixels. We can say that the histogram of the calibration region is part of the histogram of the digital image of the entire banknote, since all the pixels from each cell of the histogram of the calibration region are also included in the cell with the same number in the histogram of the digital image. Therefore, when there is a general darkening of the paper of the banknote being checked or a general abrasion of the ink layer, the characteristic level reflects this process not only on average over the banknote surface, but also in each individual calibration zone.

Consider this relationship in more detail. The characteristic level plays the role of a starting point from which the limiting binarization levels of individual calibration regions are counted. In each of the regions, the limiting level of binarization differs from the characteristic level in accordance with a coefficient that depends on the banknote pattern in this region. So, for unsealed banknote fields, the coefficient will be slightly higher than unity, since the brightness of the banknote field is higher than the average brightness of the banknote. In areas with a high density of the paint layer, the coefficient, on the contrary, will be lower than unity. As experience shows, with a general darkening of the paper, or general abrasion of the paint layer, the characteristic level and the limiting level of binarization in each of the regions change approximately the same number of times, because of which the described coefficients vary little from banknote to banknote. Thus, using the characteristic level to proportionally change the threshold value of binarization can effectively compensate for the overall darkening of the paper or the general abrasion of the ink layer.

In one implementation, a partition of the monitored zone of the banknote is used into adjacent rectangular groups of pixels of the same size, and each group is assigned a separate element of the binarization parameter map. The threshold binarization value for each pixel in the digital image of the banknote being checked is calculated as the product of the characteristic level of the digital image of the banknote being checked by the value of the binarization parameter card element corresponding to the group of the pixel in question. The threshold value for each pixel depends on a single map element of the binarization parameter, and can be calculated once for all pixels in the group. Such an implementation is distinguished by both quick calculation of the binarization threshold value and the simplicity of tuning map elements during training. The binarization parameter, presented as map elements, is a dimensionless coefficient. The dimensions of the rectangle forming the group of pixels should be selected somewhat larger than the geometric error of the scanning process in order to reduce the sensitivity of the method to these errors.

It is important to note that the number of pixels, pixel brightness levels, and threshold binarization values are specified by integers, which does not allow us to solve the problems of calculating the characteristic level and the limiting binarization level of the calibration region exactly. The presence of corresponding natural errors is implied in this description, but it is also assumed that these errors are small and do not significantly affect the result of the method. It should be borne in mind that errors increase with a decrease in the size of the calibration region, which is done to limit the size of spurious clusters. In particular, the set of possible values of the limiting level of binarization, as can be seen from its definition, coincides with the set of pixel values in the calibration region. If the region contains only a few pixels, then the limit level of binarization is calculated very roughly, which may impair the accuracy of the method.

It should be noted that, when processing banknotes, in one bundle there can be banknotes of various currencies, denominations and orientations. In a counting and sorting machine, a stack of banknotes is sequentially scanned, note by note. The digital image of a banknote obtained as a result of scanning, as a rule, has some bias. To compensate for skew, the digital image of the banknote is aligned with the coordinate axes. To determine the currency, denomination and orientation of the banknote, after scanning, usually use special methods based on pattern recognition. Alignment and pattern recognition methods are well known in the art.

When, in this description, it is said about obtaining a digital image of a banknote being checked of a certain type and orientation, this involves the steps of physically scanning the banknote, aligning the digital image to a horizontal location, as well as determining the currency, denomination and orientation. The steps of the method followed by obtaining a digital image are based on the type and orientation of the banknote that was determined when the digital image was received. Namely, a binarization parameter map is used, which corresponds to a certain type and orientation.

If a banknote processing machine involves feeding banknotes of a different type and / or different orientation, then this feature of the machine can be described using a set of several valid combinations of type and orientation. Since the map of the binarization parameter is predefined for a certain type and orientation of banknotes, for use in such a machine it is necessary to store a separate map of the binarization parameter corresponding to each valid combination. In accordance with the claimed method, a map of the binarization parameter for each such combination should be formed in advance by training using a set of images of the corresponding type and orientation.

It is possible to avoid the need to store different binarization parameter cards corresponding to different orientations of the same type of banknote. If you scan a banknote simultaneously from two sides and get digital images corresponding to these sides, then, using the conversion of digital images, you can get such a digital image that corresponds to a single, predetermined orientation. Such transformations are well known to specialists: they include the rearrangement of digital images obtained for the first and second sides of the banknote, rotation and mirroring. If, when receiving a digital image of a banknote, convert it to a predetermined orientation, then for any of the four possible orientations of a banknote of a certain type, only one card will be needed that corresponds to a predetermined orientation.

When processing banknotes, as a rule, it is required to check the banknote for the presence of a foreign image on each of its sides. This can be done, for example, by sequentially performing verification using the inventive method, first for a digital image of the first side of the banknote, and then for a digital image of the second side of the banknote. Exclusively in the case when the verification of both digital images shows the absence of extraneous images on them, we can conclude that there are no extraneous images on the banknote. It is convenient to use the conversion of a digital image to a given orientation in conjunction with checking both sides of a banknote, since in either case it is necessary to scan a banknote from both sides and similar conversions of digital images of both sides of a banknote should be carried out.

In FIG. 1 shows the location of the controlled area on the banknote, the division of this area into groups of pixels and calibration areas. FIG. 2 illustrates the binarization process of a monitored banknote zone based on a binarization parameter map. The controlled area on the initial banknote image (Fig. 2A), the conditional representation of the values of the binarization parameter card elements in the form of gray gradations (Fig. 2B), and the estimated image (Fig. 2C) are shown on the example of Russian rubles banknotes with a face value of 50 rubles of 2004 modification in facial orientation.

The block diagram of FIG. 3 shows a sequence of steps during binarization of a controlled banknote zone.

The flowcharts of FIG. 4, 5 show the sequence of steps in learning.

The block diagram of FIG. 6 shows the steps in checking a digital image of a banknote for the presence of extraneous images.

The flowcharts of FIG. 7, 8 show the steps in processing a banknote in order to check it for the presence of extraneous images.

The practical implementation of the claimed method described below, given as an example, is intended to be performed using a counting and sorting machine, which includes a banknote mechanism that enables the movement of banknotes from the supply pocket into the receiving pockets; a sensor system that allows you to scan a digital image of a banknote as it moves in the banknote mechanism; and embedded computing system. The computing system, in turn, contains a processor, memory, user interface, as well as peripheral devices and electrical circuits for connecting sensors and drives of the banknote mechanism. To ensure the convenience of training, the computing system can be additionally connected to a personal computer, to which digital images of banknotes can be transmitted, and from which training results can be downloaded in the form of a generated binarization parameter map. However, the use of a personal computer is not strictly necessary, since modern counting and sorting machines have powerful computing systems that allow you to perform all the steps of the method. In a practical implementation of the claimed method, a banknote is scanned from both sides in reflected white radiation. The digital image captured by the image sensor on one side of the banknote path, we will arbitrarily denote as image A, and the image taken on the opposite side of the path, as image B. Both digital image A and digital image B are received with a resolution of one pixel per millimeter . As shown in FIG. 1, the size of the banknote control zone 1 is selected in multiples of 4 millimeters vertically and horizontally, and the center of the control zone is positioned in the center of the digital image of the banknote. Small fields, between 2 and 4 pixels, are left between the control zone 1 and the circuit 2 of the nominal size banknote in order to compensate for the spread in the size of the digital banknote images due to the tolerances of the printing process and scanning errors.

Groups of 3 pixels for constructing a map of the binarization parameter are selected in the form of squares of four by four pixels, arranged in the order of rows and columns close to each other. Each such group of 3 pixels corresponds to one element of the map of the binarization parameter.

All calibration areas 6 and 7 are eight by eight pixel squares and their borders are aligned with the boundaries of groups 3. Each calibration area completely covers four groups of 3 pixels. In the future, for simplicity of presentation, we will talk about the fact that the calibration area covers four elements of the binarization parameter map, which, in turn, correspond to the four groups of 3 pixels mentioned above.

The set of calibration regions consists of the first subset consisting of regions 6 and the second subset consisting of regions 7. Regions 7 of the second subset in the figure are shown by a thick line. The regions 6 of the first subset overlap with the regions 7 of the second subset intersecting with them by four pixels in height and four pixels in width.

The method is designed to search for extraneous images on certain types and orientations of banknotes. The combination of the type and orientation of the banknote, for which the operation of the method is provided, constitute a set of valid combinations. The training is conducted separately for each combination of the type and orientation of the banknote from the set of valid combinations.

Here we describe the process of binarization of a digital image, which is illustrated by the flowchart in FIG. 3. When receiving a digital image, it was aligned and recognized, as a result of which its type and orientation were determined. As a result of binarization of the controlled area of the digital image, an evaluation image is created in which each pixel corresponds to a pixel of the control area of the digital image and can have one of two values: black pixel or white pixel.

First of all, the characteristic level of the digital image is calculated (step 101). Further, for each binarization parameter card element, the corresponding binarization threshold value corresponding to it is calculated and stored in the computer system memory by multiplying this element by the characteristic level of the digital banknote image that needs to be binarized (step 102). Next, for each pixel, the binarization parameter map element corresponding to it is found, and the binarization threshold value related thereto is taken as an individual threshold value for the pixel (step 104). At this step, a binarization parameter map is used, obtained in advance for those specific types and orientations to which the digital image belongs. The binarization parameter map is stored in the non-volatile memory of the computing system. The coordinates of the map element of the binarization parameter, both horizontally and vertically, are obtained by integer division of the pixel coordinates by four. Then take the value of this pixel and compare it with the individual threshold value of binarization (step 105). If the pixel value is greater than the individual binarization threshold value, then a white pixel is recorded as a result in the corresponding place of the evaluation image (step 106). If the pixel value is less than or equal to the individual threshold value, then record the black level (step 107).

Thus, the individual binarization threshold value for each pixel is a product of the characteristic level of the digital image of the banknote and the card element of the binarization parameter.

The statistical distribution of brightness in a digital image is taken into account due to the characteristic level of this digital image.

To conduct training for a certain combination of banknote type and facial orientation, use a set of two hundred banknotes of a certain type, suitable for circulation, and scan them using a banknote processing machine in the same facial orientation. For training, use the image A. Each pixel in the digital image is assigned a brightness value in the range from 0 to 255. The resulting digital images are aligned with the coordinate axes to compensate for the skew that occurs during scanning. The result is a training set containing two hundred digital images. Such a number of digital images in the training set reflects the variability of the results of the printing process and the change in the optical characteristics of banknotes as they are circulated. This amount, as a rule, is enough to conduct training, but it, if necessary, can be increased.

The purpose of the training is to fill out a binarization parameter card for its subsequent use in binarizing a digital image of a banknote to be verified. During the training, as will be shown below, binarization itself is not carried out, and the values of the binarization parameter map elements are set as a result of manipulations with the statistical characteristics of the image and calibration areas. The result of binarization of a digital image is directly determined by individual threshold binarization values of pixels, which, in turn, are determined by the map of the binarization parameter and the characteristic level of this digital image. To analyze the impact of training on the result of binarization, in our discussion we will consider how these manipulations affect the individual thresholds of binarization obtained.

The target value of the proportion of black pixels is selected high enough, but so as to provide an acceptable low level of probability of a false decision about the presence of an extraneous image for banknotes of a predetermined type and orientation that obviously does not contain extraneous images. This fraction is selected based on which detection criteria are to be used. In the described implementation, a criterion is used to search for round spots using a sliding square window, the size of which is equal to the size of the smallest spot to be detected. The criterion is triggered when the number of black pixels is detected inside the window that exceeds a predefined limit value. For a round spot, the limit value is set equal to 0.75-0.8 of the total number of pixels in the window. It has been experimentally shown that for such a criterion the best results are obtained by the target value of the proportion of black pixels, equal to 20%. This value is a good starting point. The learning process, which will be described later, provides, if necessary, the correction of the target value of the proportion of black pixels.

Training for digital banknote images of a particular combination of type and orientation is shown in FIG. 4A. First of all, all the elements of the binarization parameter map are set to a knownly large initial value, for example, equal to 10000 (step 201). Such a value, after multiplying by the characteristic level of any real banknote, will give a binarization threshold value exceeding the maximum pixel value in the digital image equal to 255. As a result, binarization of the digital image of any banknote using the initial values of the binarization parameter map elements will give an estimated image consisting of black pixels alone.

Further training is carried out sequentially using each of the digital images in the training set.

For each of the training digital images, a histogram is constructed and the characteristic level is calculated from it (step 203). To this end, find the level value at which the proportion of pixels with this and lower values in the digital image represents the target value of the proportion of black pixels of black pixels. If the exact target value of the fraction cannot be achieved, then choose a level that provides the best approximation to the target value of the proportion of black pixels, without exceeding it. Like pixel values, the characteristic level can take values from 0 to 255, but, for real digital images that have pixels with different values, the extreme values of 0 and 255 are not reached.

On each of the digital images, they successively pass through all the calibration areas, and for each of them the ultimate level of binarization is found (step 205). For this, a histogram of the pixels of the calibration region is constructed. The histogram is used to find the value of the level at which pixels with this and lower values together constitute the target proportion of black pixels in the calibration region. If the exact target value of the fraction cannot be achieved, then choose a level that provides the best approximation to the target value of the proportion of black pixels.

Next, divide the limit level of binarization of the considered area by the characteristic level of the digital image of the banknote (step 206). The result is a dimensionless binarization parameter for the calibration region. Then, it is found which exactly four elements of the map of the binarization parameter cover the region under consideration (step 207). Using the region binarization parameter as the upper limit, the values of these map elements are limited. Namely, if the found binarization parameter is less than the value of the card element, then this element is assigned the value of the found binarization parameter (step 208). As a result of this limitation, the individual binarization threshold value corresponding to each pixel in the calibration region is limited by the limiting binarization level of the region under consideration. The sequence of steps 205-208 is repeated for all remaining calibration areas and all remaining digital images from the training set, after which the training is completed. The result of the training is a binarization parameter map, which is stored in the non-volatile memory of the computer system for subsequent use in banknote control (step 209).

In accordance with the described sequence of steps, after training, for any calibration region of any digital image from the training set, the individual threshold binarization value of each of the pixels in the region is either equal to or lower than the limit level of binarization of the region. Let us consider in more detail, due to which this result is obtained. It is necessary to take into account that when we talk about a decrease in individual threshold binarization values of pixels, in fact, a corresponding decrease in the elements of the map of the binarization parameter is made, and a decrease in individual threshold values is its direct consequence.

At the very beginning of training, when considering the first area on the first digital image from the training set (when steps 203-208 are performed for the first time), individual threshold values for all pixels in this area are reduced to the limit of binarization defined for this area. Upon further consideration of other calibration areas, for each of them, their own maximum level of binarization is calculated. When any of these regions overlaps with the first region, and its limiting binarization level is less than the limiting binarization level of the first region, then at step 208, at the overlap, the individual binarization thresholds will be further reduced from the limiting binarization level of the first region. Thus, when processing a digital image of one banknote, pixels may appear in the calibration region with a threshold binarization value lower than the limiting binarization level of this region, which is due to a lower limiting level calculated for the other region with which it overlaps.

When switching to the second and subsequent digital images, the value of the characteristic level and the limiting level of binarization for each of the areas are re-calculated (step 203). Accordingly, individual binarization thresholds for each pixel are now determined based on previously obtained binarization parameter map elements and a new characteristic level. If the new limit binarization level for the region is lower than the individual threshold pixel value in this region, then the value of the corresponding element of the binarization map is further reduced (step 208) to reduce the said individual threshold value to the limit value of the binarization. For those pixels whose individual binarization threshold value already has a value below the newly calculated limiting level of binarization, further reduction is not performed. Thus, a decrease in the binarization threshold value of each pixel occurs if necessary, and to the extent that it is guaranteed to avoid exceeding the proportion of black pixels over a given target value of this fraction in each of the overlapping regions and on all images from the training set.

At step 209, the resulting binarization parameter map is stored in the non-volatile memory of the machine.

We consider an alternative variant of partitioning into gauge domains. Alternatively, the square calibration areas in the set are four by four pixels in size and arranged in rows and columns, close to each other and without overlap. Each region corresponds to one and only one element of the map of the binarization parameter. In this case, during the training (see Fig. 4B), up to and including step 206 of calculating the region binarization parameter, they act similarly to the main version. However, then, in step 221, a single element of the binarization parameter map is determined, and in step 222, it is limited to the value of the region binarization parameter. Therefore, a decrease in the value of the binarization parameter map element is due only to the appearance of a lower binarization parameter for the corresponding region in some digital image from the training set. That is, with an alternative partitioning, there is no effect of several overlapping regions on one element of the map of the binarization parameter.

An alternative has its advantages and disadvantages. Due to the reduced size of the calibration region, it allows for more detailed consideration in the map of the binarization parameter small elements of the normal banknote pattern. Due to this, the high-contrast dark elements of the normal pattern are less transmitted to the evaluation image. Therefore, the possibility of forming clusters that can cause a false decision to detect an extraneous image is reduced. In addition, the alternative option allows to reduce the values of map elements to a lesser extent, since each area affects only one map element, and not four. This allows you to increase the sensitivity of the method to the detection of low-contrast extraneous images.

The disadvantage of the alternative option is associated with the features of the histogram of pixels of the calibration region, which is built in step 205 to determine the maximum level of binarization. Since there are only 16 pixels in the region, non-zero values in its histogram will have no more than 16 levels out of 256. Therefore, the random error in determining the limit level from such a histogram will be approximately equal to the average distance between neighboring levels, that is, about 16. This is very a large value, which can randomly significantly distort one or another element of the map of the binarization parameter and reduce the reliability of detection of extraneous images. In the main version of the partition, the calibration region contains 64 pixels, therefore, the random error is much smaller and approximately equal to 4.

The choice between the main or alternative splitting option must be done on an experimental basis. For one or another type of banknote, due to differences in the features of the picture, this or that variant of splitting may be preferable. It is important that the choice of the partition option does not affect the dimension of the resulting map of the binarization parameter. Due to this, the map of the binarization parameter, regardless of the split option used, can then be used in exactly the same way to detect extraneous images.

In general, we can say that for any variant of partitioning, the map of the binarization parameter, after training, contains a generalized spatial representation of the brightness of the zone being tested for a certain type and orientation of the banknote. Such a representation takes into account the spatial distribution of the brightness of all copies of banknotes from the training set with a decrease in the influence of general darkening or lightening of each individual banknote on this representation. Turning to FIG. 2B, where the map of the binarization parameter is visualized using gray gradations, you can see its resemblance to the controlled area of the banknote reproduced in FIG. 2A. In FIG. 2B, darker gray tones correspond to lower values of map elements, and lighter - higher. The arrangement of the dark and light areas in FIG. 2B coincides with the location of the dark and light areas of the controlled zone of the banknote. At the location of the background image 8 as a dark spot in FIG. 2B, the binarization parameter map does not have a blackout, since there is no spot on normal banknotes at this point.

You may also notice that the map image of the binarization parameter of FIG. 2B has individual local dimming that is not present in the normal bill image in FIG. 2A, such as region 13. These blackouts occur during training when an increased optical density appears on any of the normal banknotes of the training set in this place. Such an increase in optical density can occur due to technological deviations of the printing process or during acceptable wear, which has not yet reached the level of decay. So, the darkened area 13 arose due to the only training copy of a normal banknote, on which in this place, as a result of the inconsistency of the printing process, an increased ink transfer from the printing form appeared. Thus, the map of the binarization parameter, although resembling an individual image of a normal banknote, also takes into account various local differences between different copies of normal banknotes. Obviously, local dimming on the map of the binarization parameter worsens the ability to detect extraneous images in these places, but reduces the likelihood of false rejection of normal banknotes. Thus, the sensitivity to the detection of extraneous images and the likelihood of a false decision to detect an extraneous image should be selected on the basis of an acceptable compromise. The greater the local differences between banknotes in the training set, the lower the sensitivity to extraneous images, but the lower the likelihood of false rejection

Complete training for all valid combinations of the type and orientation of the banknote is performed sequentially (step 230), combination by combination, as shown in FIG. 5. Step 231 combines the steps 201-209 described previously. Then, using the training set of digital images, the probability of false positives of the detection criterion is checked (step 232). To do this, verify each digital image in accordance with steps 301-308, as described below, and determine the percentage of digital images in the training set on which an extraneous image was detected. In this case, use the binarization parameter map obtained at step 231. Since there are no extraneous images on the banknotes of the training set, the percentage found is the probability of a false decision about the presence of an extraneous image. At step 233, the percentage found is compared with an acceptable level of probability known in advance based on the requirements of the central bank for processing banknotes of a certain type. The found percentage should be in the range from 1/3 of the acceptable level to an acceptable level of probability of false positives for the detection criterion. For example, if the acceptable level is 3%, then the percentage of digital images in the training set on which an extraneous image was detected should be within 1-3%. If the percentage found is too large and exceeds the acceptable level of probability, then it is necessary to slightly reduce the target value of the proportion of black pixels, and repeat the test in steps 231-232. If the percentage value found is too small, there is a reserve for increasing the sensitivity of the method by increasing the value of the map elements of the binarization parameter and the corresponding individual binarization thresholds. This reserve needs to be used, for which you need to slightly increase the target value of the proportion of black pixels, and repeat the check in steps 231-232. As a result of this iterative adjustment of the target value of the proportion of black pixels (steps 233, 231, 232), performed if necessary, there is a decrease or increase in the elements of the map of the binarization parameter. This allows us to bring the values of the map of the binarization parameter even closer to optimal.

The acceptable interval of the found percentage of false decisions about detecting an extraneous image, set from 1/3 of an acceptable level to an acceptable level of probability of false positives for the detection criterion, and used to iteratively fine-tune the target value of the proportion of black pixels, was determined empirically and its lower boundary can be changed if necessary. An increase in the lower boundary improves the sensitivity of the method, but requires an increase in the number of banknotes in the training set for more precise adjustment of the binarization parameter map elements. The lower border can be reduced to zero, in this case, the target value of the proportion of black pixels will not increase. This option can be recommended for a small available training sample size, when a certain margin of binarization thresholds is required so as not to cause frequent false positives of the detection criterion on normal banknotes, which differ significantly from banknotes included in the training sample.

When checking the banknote for the presence of extraneous images, checking its digital image is carried out, as shown in FIG. 6. First, a digital image is obtained and the type and orientation of the banknote is determined (step 301). Next, a digital image is binarized to obtain an estimated image (step 302) using a binarization parameter map stored in the non-volatile memory of the machine. This step consists of a sequence of individual steps 101-107, described in detail previously. A map of the binarization parameter for a certain combination of the type and orientation of the banknote stored in the non-volatile memory of the machine was obtained earlier during the training (steps 201-209).

Then, they begin to search for an extraneous image in the evaluation image. Set the test window to its initial position (step 303) and check the number of black pixels that fall within the window (step 304). If the number of black pixels exceeds the set limit value (step 305), then a conclusion is drawn about the presence of an extraneous image. If the number of black pixels does not exceed the set limit value, move the window to the next position (step 307) and again check the number of black pixels (steps 304, 305). Repeat steps 304-307 until you check all the positions of the test window or until you find that you have exceeded the set limit value. If for all positions of the test window, the set limit value for the number of black pixels has never been exceeded, then it is concluded that there is no extraneous image (step 307).

Turning to FIG. 2 and trace the verification of the digital image of the banknote using a specific example. As shown in FIG. 2A, within the control zone 1 on the banknote, there is an extraneous image in the form of a black spot 8. When binarizing the control zone 1 using the binarization parameter map shown in FIG. 1B, the evaluation image shown in FIG. 2C (step 302). In the part of the estimated image where the black spot 8 is absent, the black pixels are randomly scattered singly or in the form of small clusters, and their average density does not exceed the target value. Black spot 8 is transferred to the evaluation image in the form of a large cluster of black pixels 9. During the test, a sliding check window 11, passing through all possible positions on the banknote from left to right and from top to bottom (steps 303-305, 307), in the position shown in FIG. 2C, an excess of the set limit value for the number of black pixels equal to 20 is detected (step 306). Accordingly, it is concluded that an extraneous image has been detected, and further verification of the evaluation image by the verification window is terminated.

More sophisticated methods can be used to check for an extraneous image within the inspection window. For example, it can be analyzed not the number of black pixels in general within the window, but the number of black pixels in the largest cluster of black pixels in the window. Such an analysis makes it possible to exclude from consideration small clusters and single black pixels having noise origin and not related to an extraneous image. In addition to this, the ratio of the border length to the area of such a cluster can be analyzed, which allows more confidently identifying thin-line patterns. The choice of this or that method is determined both by the specific requirements for the search for an extraneous image, and by the availability of sufficient computing power to implement the method with the necessary speed.

We describe the sequence of checking banknotes in a counting and sorting machine (Fig. 7). The machine sequentially processes banknotes located in the feed pocket in the form of a bundle. Each banknote is mechanically separated from the bundle and begins to move along the banknote path. During its movement through the tract, where the sensors are installed, scanning is performed with digital images of the banknote on both sides (step 401). When scanning, various types of banknote illumination can be used, in accordance with what information is needed to recognize its currency, denomination and orientation. Namely, both monochromatic radiation sources and a combination of several wavelengths — for example, forming white light — can be used. Radiation can be supplied both from the location of the photodetector, and from the opposite side of the banknote. In the first case, a digital image in the reflected radiation will be obtained, and in the second, a digital image in the transmitted radiation. For the detection of extraneous images, the most informative are digital images obtained in reflected white light.

To compensate for the distortion of the banknote when moving through the sensor location, at step 402, the digital images are aligned with respect to the coordinate axes. To do this, by a sharp change in the brightness of the pixels, the banknote borders are located in a digital image, its center and skew angle are calculated, and then the image is rotated relative to the center by an angle opposite to the skew angle. In this case, the center of the image is shifted to a given point. As a result, the edges of the banknote are parallel to the coordinate axes, and the center of the image of the banknote is located at a given point.

Then, in step 403, the currency, face value and orientation of the banknote are recognized. To do this, use one of the methods for recognizing digital images, well known to specialists. For example, a correlation comparison can be applied with pre-prepared reference digital images of all known types and orientations of banknotes that this device processes. If the maximum correlation found with one or more of the reference digital images exceeds a predetermined threshold, then it is concluded that the banknote refers to a combination of type and orientation corresponding to the reference digital image with which the highest correlation was obtained. If the correlation with any of the reference digital images does not exceed a predetermined threshold, a decision is made (step 411) that the banknote is not recognized and should be rejected. An unrecognized banknote is not further checked for extraneous images.

If the banknote was recognized, then its digital image is then brought back to the specified orientation (step 404) by forming the first digital image. A banknote can have only four orientations, which are combinations of the side of the banknote and its rotation: face unverted, face inverted, back not inverted, back inverted. As the specified orientation, the selected face is not inverted. When the banknote already has a predetermined orientation, the image A is taken as the first digital image. If the recognized banknote has an inverted front orientation, then the digital image A is taken as the first digital image and this image is rotated 180 degrees around the center. If the recognized banknote has a back non-upside-down orientation, then the image B is taken as the first digital image. For the back upside-down orientation, the digital image B is additionally rotated around the center. As a result of these manipulations, the first digital image always corresponds to the front non-upside of the banknote.

At step 405, the first digital image is checked for extraneous images by performing steps 302-308. According to the result of the check, a conclusion is made about the presence of an extraneous image (step 409) or its absence (step 410). This conclusion is used to control the electromechanical redirection elements in the banknote mechanism. The redirecting elements are controlled from the computer system of the machine and allow you to send a banknote moving along the path into one of the receiving pockets. If the banknote was not recognized or if an extraneous image was found on it, then it is sent to the receiving pocket for the rejected banknotes. If no extraneous image was found on the recognized banknote, then it is sent to the pocket for valid banknotes. Thus, in the sequential processing of all banknotes in the supply pocket of the machine, they are sorted into suitable and unfit. Unrecognized banknotes as well as recognized banknotes with extraneous images are considered unfit.

The banknote sorting implementation described above checks only the front side of the banknote for the presence of extraneous images. This implementation has a certain practical value, for example, for those types of currencies, for which it is especially important to make sure that there are no extraneous images, for example, in the portrait of the monarch. However, for most currencies, a full check of a banknote requires control of its back as well.

Therefore, the implementation of checking both sides of a banknote has the greatest practical value. For a complete check of banknotes for a single pass through the machine, it is necessary to supplement the described implementation with additional steps for checking the back side, as shown in FIG. 8.

It is necessary, at the training stage, to form and save in the non-volatile memory of the machine a second map of the binarization parameter corresponding to the image of the back side. The training is carried out in exactly the same way as for the front side, using the scan result of the same training set of banknotes in a given orientation. However, to form a second binarization map, instead of digital image A, digital image B is used.

When checking the banknote, in step 412, together with the first digital image, a second digital image is also formed, which corresponds to the back non-upside of the banknote. When the banknote already has a predetermined orientation, digital image B is taken as the second digital image. If the recognized banknote has an inverted front orientation, then digital image B is taken as the second digital image and this image is rotated 180 degrees around the center. If the recognized banknote has a rear non-upside-down orientation, then the digital image A is taken as the second digital image. For the back upside-down orientation, the digital image A is rotated around the center.

An additional check of the presence of an extraneous image on the back side is carried out at step 407. At this step, the verification is carried out in steps 302-308, but a second digital image and a second map of the binarization parameter are used for verification. The decision on the results of the verification is made by performing steps 406 and 408. If there is no extraneous image on both the first and second digital images, then a conclusion is made that there is no extraneous image on the banknote (step 410). If an extraneous image is found on at least one of these digital images, then a conclusion is drawn about the presence of an extraneous image on the banknote (step 410). This conclusion is used to direct the banknote into the corresponding receiving pocket of the machine, in the same way as in the previously described implementation.

Next, the formation of false clusters of black dots in the evaluation image that can cause false detection of an extraneous image on a normal banknote will be considered in more detail (see Fig. 9 - Fig. 12). In contrast to the cluster 9 generated by this extraneous image 8, false clusters can occur in the estimated image of a normal banknote in the places of dark and contrasting elements of the picture.

We assume that the target value of the proportion of black pixels is 12.5%, and when searching for an extraneous image in the evaluation image, no more than 20 pixels are allowed in the sliding window 11 of 10 by 10 pixels. First, consider the basic partitioning method, which uses overlapping gauge areas. Let the adjacent regions of the banknote 6A, 6B, 6C, 6D from the first subset be arranged in such a way relative to the banknote pattern that the center of the contrasting dark image element 10 falls on their common angle (Fig. 9). Region 7 of the second subset is positioned so that it overlaps regions 6A - 6D and the dark image element 10 is located in the center of region 7. During training on the first banknote, first the binarization parameter map elements are set in the regions of the first subset 6A - 6D, and then later on the regions of the second subset, including region 7. At the time the processing of the regions of the first subset is completed, the values of the binarization parameter map elements for pixel groups 3A - 3Q will be set such that in the estimated image When you press the first training banknote in each of the areas 6A - 6D, you get 8 black pixels, which is exactly 12.5% of the total number of pixels in the area. Note that the dark image element 10, like any element of the scanned image, has pixels with different intensities; therefore, with a gradual increase in the binarization threshold value, the number of black pixels gradually increases, and with a gradual decrease in the threshold value, it gradually decreases. For definiteness, we will conditionally assume that the value of the elements of the binarization parameter map for pixel groups 3A - 3Q is set to 0.91 (Fig. 10). Note that these black pixels merge into a common false cluster 12A, consisting of 32 pixels. When passing through a sliding window 11, this cluster will be completely placed in the window and will exceed the allowable threshold of 20 pixels. Thus, false detection of an extraneous image in the normal banknote pattern will occur. The selected arrangement of the dark element 10 most stimulates the formation of a false cluster, since this element is equally distributed over the maximum possible number of calibration areas - four calibration areas 6A - 6D.

However, with continued training in the second subset of calibration areas, the values of the map elements of the binarization parameter will be corrected for groups 3F, 3G, 3J, 3K. As a result of a corresponding decrease in the binarization threshold value, the proportion of black pixels in region 7 will not exceed the target value of 12.5%. For this, the value of the card elements for groups 3F, 3G, 3J, 3K will have to be reduced to a value of 0.47 (conditional value for illustrative purposes, see Fig. 11). As a result, during binarization, the number of black pixels in areas 6A - 6D will decrease to 2 pixels each. The false cluster 12B in place of the dark element 10 will decrease to 8 pixels, which just corresponds to the target value of the proportion of black pixels in region 7. When passing through the sliding window 11, this cluster will not exceed the allowable limit of black pixels equal to 20, which eliminates the possibility of false detection of extraneous Images.

Although we considered binarization of only the first banknote, while continuing to study on subsequent banknotes from the training set, the mechanism for generating and reducing the size of false clusters will be as described previously, and the map of the binarization parameter will be gradually refined. The formation of false clusters will obey the same laws and during the check for the presence of an extraneous image when processing banknotes in a counting and sorting machine. This example illustrates how the location of the calibration areas overlapping each other avoids the formation of large false clusters of black pixels in the evaluation image of a normal banknote.

In the case of an alternative partition, the calibration areas coincide with the pixel groups 3A - 3Q (see Fig. 12). The dark image element 10 falls into the regions 3F, 3G, 3J, 3K, for each of which during training the value of the binarization parameter map element will be set equal to 0.26 and providing 2 pixels in each of the regions. Note that the number of 2 pixels in each of the areas 3F, 3G, 3J, 3K just corresponds to the target value of 12.5% of the proportion of black pixels in the area. As in the case of the main variant of the partition, the entire false cluster 12C will be only 8 pixels and will not be detected by a sliding window as a false sign of an extraneous image.

Thus, both the main and alternative partitioning methods solve the problem of suppressing large false clusters in the estimated image of a normal banknote. In the main version, the estimation of the limiting level of binarization is carried out over a large calibration region measuring 8 by 8 pixels, which to a lesser extent adapts the binarization threshold map to small contrasting elements in a normal banknote image. An alternative option, due to the smaller size of the calibration area, better reflects such contrasting elements in the map of the binarization parameter and allows more reliably to transfer a low-contrast background image, which is against the background of these elements, in the evaluation image. However, it should be borne in mind that in the alternative variant of partitioning, due to the small number of pixels in the region, the value of the binarization parameter map element will be set with a large random error, which may impair the reliability of detection of extraneous images.

The used options for dividing the controlled zone into calibration areas, and strategies for reducing the values of the map elements of the binarization parameter, do not exhaust all the possibilities. If necessary, more complex methods can be applied, corresponding to the claims of the claimed invention.

Claims (52)

1. A method for detecting extraneous images on a banknote, in which a digital image of a banknote to be checked for the presence of extraneous images is obtained, and its correspondence to one of the predefined types and orientations is determined, and then, using the shape and location of the controlled area on the banknote surface corresponding to this type and orientation of banknotes, an estimated image is formed by binarizing the controlled area, and the individual binarization threshold value for each pixel of the controlled area is calculated based on the map of the binarization parameter predefined for a certain type and orientation of banknotes, and is associated with the elements of this map with a known functional dependence, moreover, one element of the binarization parameter map determines the threshold binarization value for a group of at least two adjacent pixels, and the known functional dependence is established taking into account the statistical distribution of pixel luminances in the digital image of the banknote to be checked in order to reduce the effect of the brightness spread on the result of binarization on various banknote copies, and then, an extraneous image detection criterion based on the analysis of black pixels in that evaluation image is applied to the evaluation image, and based on this criterion, a decision is made on the presence of an extraneous image in the controlled area of the digital image of the banknote to be verified, moreover, the elements of the binarization parameter map are selected in such a way as to increase the corresponding individual binarization thresholds and at the same time provide an acceptable low level of probability of a false decision about the presence of an extraneous image when performing the indicated sequence of actions for banknotes of a predetermined type and orientation that do not contain extraneous images . 2. The method according to claim 1, in which a map of the binarization parameter corresponding to a certain type and orientation of banknotes is obtained by pre-processing a set of genuine banknotes that do not contain extraneous images, during which they receive a set of training digital images of a set of genuine banknotes that do not contain extraneous images for each predetermined type and orientation, and form a map of the binarization parameter on the basis of the obtained set of training digital images, for which a set of calibration areas is set that together cover the controlled area, and also the target value of the proportion of black pixels, less than 40%, establish such initial values of the binarization parameter map elements that when binarizing each digital image from the training set, the proportion of black pixels in each calibration region exceeds the target value, and further, when considering each calibration region on each of the training digital images in the set, by changing the map elements of the binarization parameter, the binarization threshold value for individual pixels is reduced then and to the extent that it is necessary to reduce the proportion of black pixels within each calibration region to level no more than target value. 3. The method according to p. 2, in which when considering each calibration area on each of the training images calculate the limit level of binarization for the region at which the target value of the proportion of black pixels is achieved in this region, and to reduce the binarization threshold value for each pixel in this region, if the previously set binarization threshold value of a given pixel exceeds the limit level of binarization, reduce it to the limit level of binarization by changing the map elements of the binarization parameter. 4. The method according to p. 3, in which the verification of the detection criteria consists in the fact that they use a test window of a predetermined size, and for all possible positions of the test window within the bounds of the estimated image, the proportion of black pixels among all pixels falling within the bounds of the check window is counted, then this fraction is compared with a predetermined limit value; if the limit value is exceeded, a conclusion is made about the detection of an extraneous image. 5. The method according to any one of paragraphs. 1-4, in which when obtaining a digital image of a banknote of a predetermined type and orientation, the combination of the type and orientation of the banknote refers to a set of valid combinations containing at least two such combinations. 6. The method according to any one of paragraphs. 2-5, in which when defining a set of calibration areas, it is ensured that at least a portion of the pixels of the monitored area simultaneously belong to at least two calibration areas. 7. The method according to any one of paragraphs. 1-5, in which to obtain a digital image of a banknote scanning a banknote as the banknote passes through the banknote path of the banknote processing machine to obtain digital images of the banknote for both sides thereof, align the resulting digital images of the banknote with respect to the coordinate axes, recognize the type and orientation of the banknote, converting digital images of the banknote to a predetermined orientation, obtaining a first digital image corresponding to the first side of the banknote in a given orientation, and this digital image is used to check for extraneous images in accordance with the recognized type and a given orientation. 8. The method according to p. 7, in which the check for the presence of extraneous images is carried out twice, why first, when converting digital images of the banknote to a predetermined orientation, a second digital image corresponding to the second side of the banknote in a predetermined orientation is additionally obtained, and then the first digital image and then the second digital image are used first, after which a decision is made that there are no extraneous images on the banknote to be verified only if together conclusions are drawn about the absence of an extraneous image in the controlled area of both the first and second digital images of the banknote, otherwise, they decide on the presence of an extraneous image on the banknote, moreover, the binarization parameter map used to binarize the first digital image is constructed as a result of training using genuine first banknotes of a certain type and predetermined orientation, not containing extraneous images, using a set of first educational digital images. and the binarization parameter map used to binarize the second digital image is constructed as a result of training using a set of second training digital images of genuine banknotes of a certain type and given, not containing extraneous images. 9. The method of claim 6, wherein calibration areas are rectangular with constant width and constant height, and the set of these calibration areas consists of two subsets, so that these calibration areas in each subset are placed close to each other in the order of rows and columns, moreover, the calibration regions of one subset overlap with the intersecting calibration regions of the other subset half the height and half the width of one calibration region. 10. The method according to p. 6, in which calibration areas are rectangular with constant width and constant height, so that the calibration areas are arranged in row and column order, moreover, within a series of calibration areas are placed with a predefined first step, and the rows are offset relative to each other with a predetermined second step in a direction perpendicular to the direction of the rows. 11. The method according to any one of paragraphs. 2-5, in which a set of calibration areas is defined in the form of rectangles of the same size, arranged in order of rows and columns close to each other, the size of each rectangle is chosen small enough to avoid the formation on the evaluation image of a banknote that does not contain extraneous images, spurious clusters of black pixels having a size more than a predetermined one. 12. The method according to any one of paragraphs. 3-6, in which the known functional dependence is established taking into account the statistical distribution of pixel luminances in the digital image of the banknote to be verified due to the fact that for the banknote to be verified, a characteristic level is calculated at which the fraction of pixels with equal or lower values in the digital image of the indicated banknote to be verified is the target value of the proportion of black pixels predefined for this type and orientation of banknotes, and the value supplied by the functional dependence is proportional to the characteristic level. 13. The method according to p. 10, in which the controlled area is divided into rectangular adjacent groups of pixels of the same size, with each group corresponding to a separate element of the binarization map, and the binarization threshold value for each pixel in the digital image of the banknote to be checked is calculated as the product of the characteristic level of the banknote to be checked by the value of the binarization map element corresponding to the group of the pixel in question.

Images