RU2718571C1 - Method for binarisation of images of symbols on a banknote based on histogram of length of boundaries - Google Patents

Abstract

FIELD: data processing.SUBSTANCE: invention relates to a method of binarisation of images of symbols on banknotes and can be used during authentication. To a halftone image of a symbol in a portion of a banknote to be binarized, a pixel of which is characterized by a brightness level, applying a threshold procedure using a process of forming a two-level image having one of two possible brightness levels with a predetermined binarisation threshold, plotting a histogram of the boundaries of said image, for which an ordered set of increasing brightness values, which are possible in a halftone image, is set, and for all values of brightness in the specified set, starting from the value following the least, the value of the corresponding cell of the histogram of the length of boundaries is set as an increment of the length of boundaries of the two-level image. Length of boundaries of the two-level image is counted as the sum of the lengths of the segments, each of which is a boundary segment of two adjacent pixels with different brightness levels, adding a component to account for the length of those sides of the pixels, which are located on the edge of the image, analyzing said histogram of the length of boundaries to find the final threshold of binarisation, when using which the result of binarisation contains continuous lines contained in the gray-scale image of the banknote area.EFFECT: faster recognition of banknote symbols.7 cl, 6 dwg

Description

The invention relates to a method for binarizing symbol images on banknotes and can be used in authentication. A banknote is a secure document containing both pictorial and textual information. Textual information consists of characters that are letters, numbers, and some characters, such as forward and backward lines. In this description, we will not refer to textual information those characters that are artistically embedded in the drawing on the banknote and, thus, are an integral part of the pictorial information.

Part of the textual information is invariably applied to each banknote of a certain denomination and version of its issue. Such information includes, for example, an indication of the version of the issue of the banknote and the place of its issue, as well as the name of the treasurer of the national bank. In addition, each instance is marked with a unique character string, which is usually called a serial number. Controlling unchanged textual information is important for banknote authentication. However, it is more important to provide recognition of variable textual information. In modern technology for processing banknotes, it is necessary to keep track of the movement of individual copies of banknotes along chains of money circulation. Thus, for the processing of banknotes, equipment is needed that can recognize the characters on the banknote located in the serial number, as well as in the locations of invariable text information. Character recognition on banknotes is a special task, which differs from the widely used text recognition methods used in office and library practice. Documents and print media, with the exception of a small number of special types of documents (historical documents, ancient literature), have high print quality. As a rule, they are printed in black on a white background, and paper pollution and damage to the paint layer are practically absent. When scanning such documents, a contrasting and complete image of each character is obtained. The scan resolution used for documents and prints is typically 300 dpi or more.

Unlike documents and print media, banknotes are constantly in circulation, which leads to the gradual abrasion of the paint layer, staining and general pollution of the paper. As a result, the contrast and image integrity of the symbol on banknotes can be very noticeably reduced. The situation is also complicated by the fact that the serial number or unchanged text is not always printed on the white section of the banknote. In many cases, it is applied to a previously sealed area of paper. Therefore, even on new banknotes, the background of the symbol often already contains an additional image that interferes with recognition.

The quality of scanning banknotes in cash processing machines is significantly inferior to the quality of scanning conventional documents, since the banknote moves in the machine at high speed (up to 10 kilometers per hour). This leads to the need to use a low resolution scan, not exceeding 200 dpi.

Due to the high speed of movement of the banknote in the machine, high demands are placed on the speed of character recognition on the banknote. Typically, it can take a few milliseconds to recognize a serial number string. This time is approximately comparable with the recognition time of a printed word of the same length in modern methods of recognition of ordinary documents, executed on high-performance processors. However, the use of high-performance processors in most banknote processing machines is not possible for economic reasons. The same types of processors that are economically viable for use in banknote processing machines have 10 or more times lower performance than the processor performance of a workstation with an IBM PC architecture. In this regard, the computational complexity of recognizing characters on a banknote should be reduced in comparison with the recognition methods used to recognize ordinary documents.

Character recognition methods on banknotes can be divided into those that use pre-binarization of a grayscale image and those where the recognition algorithm directly refers to a grayscale image. An example of a recognition algorithm that does not require binarization is the convolutional neural network method, as well as the gradient histogram method. In binarization methods, pixels belonging to a significant object in the image are first separated from background pixels. Binarization is a method of highlighting a significant object in a grayscale image, such as a symbol, based on the brightness of the pixels. Usually, the result of binarization is presented as an image with two gradations of brightness, where the background pixels appear white and the object pixels are black. The recognition algorithm works with a binarized image. Such recognition algorithms include, for example, morphological algorithms. These algorithms analyze the binarized symbol image for the presence and location of certain formal symbol elements, such as straight lines, arcs, symbol crosses, as well as background elements in the form of “islands” and “peninsulas”.

The pixels of an object are often called the foreground pixels of the image. In the framework of this description, a symbol is considered as an object, therefore, the foreground pixels of an image are called pixels assigned to a symbol. Keep in mind that the quality of binarization can be different. If the quality is poor, the pixels of a symbol during binarization may mistakenly include pixels that, in reality, do not belong to the symbol and are part of the background or an extraneous object in the image. Similarly, pixels that actually refer to a character may be erroneously assigned to the background. As experience shows, it is impossible to create a completely error-free method of binarization, therefore, these errors, one way or another, always take place.

In general, binarized image algorithms require less computational resources and are faster than algorithms that directly access a grayscale image. Therefore, binarization is often used in banknote processing machines. At the same time, binarization, in itself, requires additional processor time. If it fails, binarization can degrade the image quality of the symbol to such an extent that its recognition becomes impossible. In such cases, the image of the symbol consists of separate divided segments with the loss of certain parts of the symbol, or, conversely, background fragments are connected to the symbol and violate the structure of its component parts. Another type of error introduced during binarization is the manifestation of isolated background fragments in the form of points and lines that are mistakenly recognized by the recognition algorithm as elements of a symbol.

Thus, the binarization method for use in banknote processing machines must have a certain set of properties. Firstly, it must be executed quickly, so as to allow the entire recognition process to fit into the short time allotted for it. Secondly, the symbol must be transmitted to the binarized image without loss of parts and line breaks. Thirdly, the binarization method should reliably block the appearance of background pixels in the binarized image as pixel symbols, despite the fact that the background itself can be a high-contrast image, and the brightness of individual background pixels can come close to the brightness of the pixel pixels.

The grayscale digital image of a banknote obtained by scanning in a banknote processing machine has some features that make the binarization task somewhat easier. When scanning, the illumination of the surface of the banknote in the banknote processing machine is very uniform, therefore, in the digital image, there is no joint change in the brightness of parts of the symbol and parts of the background, due to the uneven illumination. When binarizing a symbol image, there is no need to compensate for uneven illumination, which creates significant difficulties in solving related technical problems, such as recognition of license plates of cars.

In addition, with rare exceptions, the interval of the brightness of the pixels on the banknote in the vicinity of the symbol does not overlap with the interval of the brightness of the pixels of the symbol. In most cases, these intervals are in contact at a separating point on the brightness scale, on one side of which are the pixel pixel brightness values, and on the other are the background pixel brightness. This allows you to separate the background pixels and the pixels of the symbol using a single, so-called global, binarization threshold, which is equal to the brightness at the specified dividing point. All pixels whose brightness is less than the brightness of the binarization threshold are assigned to the symbol, and the remaining pixels are assigned to the background. The possibility of using the global binarization threshold significantly accelerates the binarization process, since the threshold must be calculated only once and then repeatedly applied to the pixels of the digital image of the banknote where the symbol is located. At the same time, the choice of the binarization threshold is a crucial task, on which, mainly, the quality of binarization of symbols on the banknote depends.

To minimize computational costs, binarization and recognition is carried out for a previously known portion of the digital image of the banknote, on which the desired symbol or group of symbols is located. As a rule, before character recognition, the digital image is analyzed to determine the currency, face value, version of the banknote issue and its orientation when passing through the banknote processing machine. The methods used in practice for such an analysis are well known in the art. According to the results of this analysis, based on the well-known graphic design of a banknote, it is possible to select a portion of a digital image for recognition of a symbol or symbols with high accuracy.

Printed characters on a banknote, as a rule, have a high optical density of the ink layer and clear boundaries. Due to the local instability of the ink transfer in the printing process, the optical density within the symbol changes, as a rule, by a small amount, not exceeding several percent. In addition, as the banknote wears out, the optical density of the ink layer becomes different at different points of the symbol. These differences are transmitted in a digital image of a banknote in the form of an inevitable spread of brightness of pixels related to a symbol.

The background on which the symbol is printed may be an unprinted surface of the paper or a printed pattern. For polymer-based banknotes, the background is often made as a solid seal with white paint. The brightness of the background pixels in a digital image is inconsistent, even if there is no printed pattern in the background. The local reflectivity of the paper surface is inconsistent and is determined by its fibrous structure. Continuous white paint also produces uneven brightness due to local unevenness in the thickness of the ink layer. An additional contribution to the brightness inhomogeneity of both the printed symbol and the background is made by the noise of the photodetector and electronic equipment used to register the digital image.

Due to the blur in the optical system of the photodetector sensor used to obtain the digital image, an uncertainty band is formed on the digital image of the banknote surrounding the lines of each symbol. The pixels in the uncertainty band have a brightness that is the sum of the brightness of the border of the symbol and the brightness of the background. It is important that the contribution of each of these brightnesses depends on a mass of random factors, such as the displacement of the surface of the banknote from the point of best focus of the optical system of the photodetector, as well as the distance between the pixel border and the nearest symbol border. Therefore, for pixels in the uncertainty band, the ratio between the contributions of the symbol brightness and the background brightness cannot be reliably determined. This creates a difficult problem to assign each pixel in the uncertainty zone to the pixels of the symbol or to the pixels of the background.

Let us consider how the binarized image changes in stages with a sequential increase in the global binarization threshold. At the first stage, at a very low value of the global threshold, the brightness of almost all the pixels in the digital image is higher than the threshold value. As a result, the binarized image consists almost exclusively of white pixels. As the threshold grows, it is higher than the darkest pixels of the symbol. In this second stage, scattered black pixels related to the symbol begin to appear in the binarized image. During a further increase in the threshold, in the third stage, disparate black pixels are combined into clusters, which are parts of a symbol separated by gaps consisting of white pixels. Further, these gaps are reduced, and at a certain threshold value, which we will call the minimum acceptable, the gaps disappear almost completely. Both in the second and in the third stage, the symbol does not appear simultaneously in the binarized image, since the pixels of the symbol have different brightness levels and begin to be transmitted to the binarized image at different threshold values.

At the moment of almost complete disappearance of gaps, with the so-called minimum acceptable threshold value, the symbol in the binarized image consists of continuous lines corresponding to the lines of the original symbol on the banknote. Possible random single gaps do not violate character recognition capabilities. The outline of the symbol in the binarized image, in general, corresponds to the original, but the thickness of the lines is significantly less.

During the fourth stage, a subsequent increase in the threshold leads to the fact that the pixels in the uncertainty band gradually change from the background to the symbol, which is accompanied by an increase in the thickness of the lines while maintaining the style. There is a threshold value, which we will call best, at which the average thickness of the lines of the symbol in the binarized image is equal to the average thickness of the lines of the original symbol. With an increase in the threshold in excess of the best, an increase in the thickness of the lines continues, while maintaining the style. There is a maximum threshold value at which the character design is still preserved, despite the fact that the thickness of the symbol lines in the binarized image exceeds the thickness of the lines of the original symbol. At the fifth stage, when the threshold is increased above the maximum allowed, some background pixels in the digital image appear below the threshold and appear in the binarized image as black pixels. As a result, the style of the character is broken and noise clusters of black pixels appear in the spaces. At the sixth stage, at very large threshold values, the noise clusters merge with the symbol lines. At the seventh stage, noise clusters and symbol lines merge with each other, forming a solid black filling of the binarized image with white inclusions. At the eighth stage, at the highest threshold value, white inclusions disappear, and the entire binarized image consists of black pixels. At the fifth through seventh stages, all background pixels do not simultaneously transition to symbol pixels, since background pixels, even in the absence of a background image, have different brightness levels and turn into symbol pixels in a binarized image at different threshold values.

It is obvious to a specialist that for guaranteed character recognition, during the binarization it is required to preserve its mark. This is possible when the binarization threshold is between the minimum allowable and maximum allowable values, which corresponds to the fourth stage of increasing the threshold.

In order to obtain the best conditions for character recognition, it is necessary that the binarization threshold is equal to the best value. Then, in the binarized image, black pixels most accurately reflect the image of the original symbol on the banknote, while preserving all the form elements of the symbol and the thickness of its lines. If the threshold between the minimum acceptable and the best values is selected, then the pixels in the uncertainty band that actually refer to the symbol may be mistakenly assigned to the background. If the threshold between the best and the maximum allowable values is selected, then, on the contrary, the pixels in the uncertainty band related to the background may be erroneously assigned to a symbol. Wrong choice of the threshold leads to thinning or, conversely, to thickening of the lines of the symbol, as well as to distortion of small shaped elements, such as serifs at the ends of the lines. In some fonts, the lines come very close to each other, forming a narrow gap. Near the maximum allowable threshold, due to an excess increase in the thickness of the lines, the lines can close, so that the gap can disappear. The listed distortions create difficulties for the recognition algorithm using a binarized image.

Thus, the choice of a global threshold is a crucial task. We can talk about a rough choice of a threshold when its value is ensured in the interval between the minimum allowable and the maximum allowable. However, for the best character recognition quality, accurate threshold selection is required when the threshold receives a value close to the best.

A well-known way of choosing a global binarization threshold is the so-called Otsu method (Otsu or Father is also written in the literature). In this method, the global threshold is determined solely on the basis of the analysis of the histogram of image brightness. The Otsu method is based on the assumption that the luminance histogram is a bimodal statistical distribution in which one mode refers to a symbol and another mode refers to a background. Under this assumption, the optimal binarization threshold should be located at the boundary of these modes. In statistical terms, the Otsu method divides the image brightness histogram by the global threshold value into two parts, called classes, so that interclass dispersion is maximal. The Otsu method uses enumeration of all brightness values in a histogram and, in the initial implementation, requires significant processor time. However, improvements are known for him that use recurrence formulas and reduce the total number of calculations. Therefore, the generally accepted implementation of the Otsu method has moderate computational complexity.

As applied to symbols on banknotes, the Otsu method successfully provides a rough choice of the threshold if the background of the symbol is bright and uniform or low contrast. In such cases, the histogram of the grayscale portion of the digital image to be recognized has an explicit bimodal character. Therefore, the global binarization threshold, found by the Otsu method, stably divides the intervals of brightness of pixels related to the symbol and related to the background. However, when the background is a high-contrast image, the bimodal nature of the histogram is violated. For such a histogram, the Otsu method very often finds an incorrect binarization threshold that lies outside the interval between the smallest and largest acceptable values. Often, the histogram is multimodal, with more than two modes. In the case of a multimodal histogram, the background may correspond to two or more modes of the histogram. In this case, the found threshold often separates the pixels of the symbol and part of the background pixels from another part of the background pixels near the point where the background modes border each other. This merges the character with part of the background pixels.

In a more complex case, the histogram has one mode, where the real dividing point, separating the intervals of the brightness of the symbol and background, is within the same mode. As a rule, this happens if the statistical distribution densities of both the symbol pixels and the background pixels in the histogram near the dividing point turn out to be quite high. In addition, it is possible that the histogram, due to the bimodal distribution of the brightness of the background pixels, is also bimodal, but the dividing point is within the same mode. For such a distribution, the Otsu method does not always guarantee even a rough choice of the threshold.

Known patent US 5956421 (publ. 09/21/1999, IPC G06K 9/38), hereinafter referred to in this description as the Tanaki method. In the Tanaki method, only the brightness histogram data is used to find the global threshold, as in the Otsu method. To search for a threshold, a sequential approximation process is used, at each step of which a new position of the working segment of the histogram levels is calculated. Initially, the working segment covers the entire histogram. At the first step, the mathematical expectation and the asymmetry parameter of the distribution on this segment are calculated. The asymmetry parameter is the ratio of the third and second central statistical moments. If the asymmetry parameter is negative, then the upper end of the segment is set to the previously calculated point of mathematical expectation. If the asymmetry parameter is positive, then the lower end of the segment is set to the previously calculated point of mathematical expectation. Next, the step is repeated. At each step, there is a decrease in the length of the segment. The process is stopped as soon as the module of the asymmetry parameter becomes less than unity. The mathematical expectation obtained in the last segment is used as the global binarization threshold. In the case of a close brightness level of the symbol and background, and also with a contrasting background, the Tanaki method finds a significantly better position of the dividing point than the Otsu method. It guarantees a rough determination of the threshold in almost any background. However, the Tanaki method has high computational complexity, since at each step it is necessary to calculate, at all levels on the working interval, not only the mathematical expectation, but also the sum of the squares and cubes of the deviation from this mathematical expectation. A well-known patent describes a hardware accelerator for implementing the Tanaki method, designed to provide acceptable performance.

Unfortunately, neither the Otsu method nor the Tanaki method allow an accurate determination of the global threshold. In general, this problem is characteristic of those methods where, exclusively, a luminance histogram is used as data for finding the global threshold. Since for the pixels in the uncertainty band the relation between the contributions of the symbol brightness and the background brightness is not known reliably, it is not possible to develop a criterion by which to divide the levels in the histogram that correspond to background pixels and symbol pixels. Differences in the brightness of neighboring background regions create additional difficulties in dividing the uncertainty band. Most often, this problem leads to an uncontrolled change in the thickness of the symbol lines in the binarized image depending on the brightness and contrast of the background and the state of the scanning mechanism at a particular moment of scanning, as well as other random factors.

Thus, the use of only a histogram of the brightness of a portion of a digital image does not make it possible to reliably find the best value of the global binarization threshold for many practical applications.

In the scientific literature, the so-called adaptive binarization methods are known, in which instead of using a global threshold, an individual threshold value is used, separately calculated for each pixel. Examples include the Otsu adaptive method, the Bradley, Niblack, and Sauvola methods. In general, the adaptive method analyzes the neighborhood for each pixel, and determines the required individual threshold from the brightness values of pixels in this neighborhood. In the adaptive Otsu method, a luminance histogram is constructed in a given neighborhood of the pixel, and using it, similarly to the usual Otsu method, a dividing point is determined. This dividing point is used as an individual threshold for the pixel. In the Bradley method, the average brightness level is calculated within the vicinity of the pixel, and then this average level is multiplied by some tuning factor. The value of the tuning factor is selected slightly less than one. As a result of multiplication, an individual threshold for the pixel is obtained. In the Niblack method, to obtain an individual threshold, the variance of brightness in the vicinity of the pixel, multiplied by the tuning factor, is subtracted from the average brightness level. In the Sauvola method, the average brightness level is multiplied by a variable coefficient, which linearly depends on the brightness variance in the vicinity of the pixel. Specialists are aware of the further development of the Sauvola and Niblack methods, using more complex functional dependencies to calculate an individual threshold.

Adaptive binarization methods take into account the state of the nearest background pixels around the symbol and, to a certain extent, reduce the influence of the brightness and contrast of the background on the allocation of the symbol pixels in the uncertainty band. The main disadvantage of all adaptive binarization methods is their increased computational complexity. Of those listed here, only the Bradley method has the computational complexity and speed acceptable for use in banknote processing machines. All other adaptive methods are many times slower than the Bradley method. In terms of the quality of separation of a high-contrast background from a symbol, the Sauvola method shows fairly good results. Unfortunately, it also turns out to be the slowest. The Bradley method resolves some of the difficulties caused by the presence of a high-contrast background. However, when using it, as a result of binarization, randomly located clusters of black pixels appear on large areas of the background, and the thickness of the symbol lines decreases at their intersections. Thus, well-known adaptive algorithms are of little use for binarizing symbols on a banknote.

Halftone images of symbols are characterized in that the pixels of the symbol have similar and low brightness values. Symbols of both European and Asian alphabets consist of lines whose length is many times greater than the thickness. Therefore, the pixels of the symbol are adjacent to each other, forming elongated clusters. These elongated clusters contain a significant number of pixels and correspond to the lines of the symbol. The presence of large elongated clusters of neighboring pixels having close and low brightness values distinguishes characters from other image elements. The background forms clusters of high brightness. Noise elements caused by the fibrous structure and surface contamination of the paper form small dark clusters. Graphic elements that do not consist of lines form dark clusters that do not have an elongated shape. The clustering properties of the pixels forming the symbols listed here are used as additional information to find the optimal global threshold. Thus, US patent US 9367899 (published on 06/14/2016, IPC G06K 9/00) is known, in which test binarization of a digital image is performed at various threshold values, and the result of binarization is analyzed. In each of the obtained binary images, clusters of black pixels are searched for, their geometric parameters are measured, and based on these parameters, clusters are assigned either to symbols or to noise elements. For each trial threshold value, the ratio between the number of noise clusters and symbol clusters is calculated. The threshold at which this ratio is minimal is considered optimal.

In accordance with the previously described steps for changing a binarized image, noise clusters are present in large numbers in this image in the second, third, and fifth stages. At the fourth stage, corresponding to a rough definition of the threshold, noise clusters are absent or their number is minimal. Thus, the optimal threshold determined in the known patent corresponds to the fourth stage.

The known method is very computationally complex due to the repeated repetition of binarization and the search and evaluation of clusters in numerous trial binary images. So, in the practical implementation of the method, 16 binary images are created, one for 16 trial threshold values. Due to the high computational complexity, it is not suitable for character recognition on banknotes. The threshold he finds is guaranteed to be in the interval between the smallest and largest acceptable values. However, this method does not accurately determine the threshold, at least in the implementations described in the patent. In this regard, he shows a result similar to that of the Tanaki method.

A number of methods are known for binarizing symbol images based on comparing the brightnesses of neighboring pixels and examining the so-called related components in the original digital image, consisting of neighboring pixels of similar brightness.

In US patent US 10049291 (publ. 02.15.2018, IPC G06K 9/34) the image is divided into related components consisting of adjacent pixels of the same brightness level. Next, build a hierarchical graph of the connection of related components to each other on the principle of touching pixels related to various related components. Then, analyze the ratio of the areas of the connected related components and get the so-called sharpness parameter. The sharpness parameter, as well as the parameters of contrast and area, make it possible to distinguish the related components that together make up the symbol from the noise and background related components. As a result, the binarization threshold is selected, which ensures the complete transmission of the symbol to the binary image. Due to the use of a hierarchical communication graph, only related components assigned to the symbol are transferred to the binary image, no noise clusters that are not associated with the symbol are not transferred. The described method provides reliable separation of the symbol from the background, but has increased computational complexity. This complexity is associated primarily with the costs of tracking chains of adjacent pixels in separate related components, constructing a hierarchical graph of the connection of related components, and computing the characteristics of the connection between related components. The more complex the communication graph is and the more connected components in the image, the longer binarization lasts. This does not allow to meet the stringent requirements for speed when processing banknotes. The aforesaid can be attributed as a whole to methods that use the construction of hierarchical graphs of related components.

Like all other methods described earlier, this method does not contain a criterion by which you can choose the threshold that best separates the pixels in the uncertainty band. The sharpness parameter allows you to determine the line, but not the thickness of the line. Thus, with binarization, the exact transfer of small shaped elements and the thickness of the lines of the original symbol is not guaranteed.

In the case of symbols on banknotes, the font with which the symbols are printed is always known, and therefore the average thickness of the symbol lines is known in advance. This makes it possible to search for the best binarization threshold, providing a previously known average symbol thickness. This allows us to eliminate or significantly reduce the influence of the zone of uncertainty on the accuracy of reproduction of the symbol in the binarized image.

To find the best threshold, one could estimate the thickness of the lines obtained by using one or another binarization threshold. The binarization threshold should be chosen as the best value at which the average line thickness in the binarized image corresponds to a previously known one. This could be done, for example, using the method described in the aforementioned US Pat. No. 3,967,899. To do this, it would be necessary to measure the thickness of the lines in the clusters on trial binarized images and compare the measured thickness with a predetermined value. The threshold at which almost all binarized image clusters have a predetermined average line thickness would be close to the best. Unfortunately, due to the computational complexity indicated earlier, such a solution is not suitable for recognizing characters on banknotes.

US patent 9367899 was selected as a prototype of the claimed invention. The technical result of the claimed invention is to increase the speed of recognition of banknote characters.

This result is achieved in a method for binarizing symbol images on a banknote, in which a grayscale image of a section of a banknote containing at least one symbol to be binarized in which each pixel is characterized by a brightness level is obtained, and a final binarization threshold is determined and a final binarization result is generated, why apply the threshold procedure using the final binarization threshold where the threshold procedure is the process of forming a two-level image in which each the second pixel corresponds to a grayscale image pixel, and has one of two possible brightness levels, assigned based on a comparison of the brightness of the corresponding grayscale image pixel with a given binarization threshold, and a histogram of the boundaries of the specified grayscale image is constructed to determine the final binarization threshold, for which an ordered set of increasing luminance values possible in a grayscale image, and, for all luminance values in a specified set, starting from the value following for the smallest, specify the value of the corresponding cell of the histogram of the length of the boundaries as an increment of the length of the boundaries of the two-level image obtained as a result of the threshold procedure, corresponding to the increment of the threshold used in the threshold procedure to the brightness value considered from the previous brightness value in the specified set, and the length of the boundaries of the two-level image is calculated as the sum of the lengths of the segments, each of which is the border of two adjacent pixels with different levels brightness, plus an additional component to take into account the length of those sides of the pixels that are on the edge of the image, and according to a specified criterion, analyze the specified histogram of the border lengths to find the final binarization threshold, when using which the binarization result contains continuous lines contained in the grayscale image of the banknote section.

Obtaining a grayscale image of a portion of a banknote containing a symbol is a typical banknote processing operation and is well known in the art. The application of a threshold procedure using a given global threshold to obtain the final result of binarization is also a typical operation used in binarization methods based on the global threshold. The novelty and inventive step of the claimed invention are provided with features characterizing how the final binarization threshold is determined in it.

In the prototype, in order to determine the final binarization threshold, a lot of trial binarizations with trial threshold values are produced, the results of trial binarizations are analyzed, and the final threshold is selected according to the analysis results. The claimed invention also considers a set of possible values for a threshold, however, to select the final threshold from them, it is not necessary to perform many test binarizations. Instead, a histogram of the length of the borders is used to analyze the quality of binarization and select the appropriate final threshold. Consider the histogram of the length of the borders in more detail.

The length of the borders in a two-level binarized image is the sum of the lengths of the segments, each of which represents the border of two adjacent pixels with different levels of brightness. If two bordering pixels in a binarized image have the same brightness value, then their boundary segment is not included in the length of the borders. The histogram of the length of the borders shows how the length of the borders changes with a stepwise increase in the binarization threshold, which is used to obtain a binarized image.

When examining two adjacent pixels of a grayscale image, there is a simple way to determine at what values of the binarization threshold the boundary segment between these pixels will be included in the length of the borders of the binarized image. For each of the boundary segments between two pixels of a grayscale image, if the binarization threshold is in the range between the brightness values of these pixels, then binarization will produce pixels with different brightness levels. In this case, their boundary segment is included in the length of the boundaries. If the binarization threshold is lower than the specified interval or higher than the specified interval, then binarization will produce pixels with the same brightness levels. In this case, their boundary segment is not included in the length of the boundaries.

The inclusion of the ends of the interval or their exclusion depends on how exactly the comparison operation is defined in a particular implementation of the threshold procedure.

When some sides of the pixel part of the symbol appear on the border of the banknote section, they must be considered in a special order. They are not boundary segments between the pixels of the image section of the banknote. To account for the contribution of these parties to the length of the borders in the claimed method provides an additional component of the length of the borders. In particular, for dark symbol pixels and light background pixels, the additional component of the length of the borders of a two-level image is the sum of the lengths of those sides of the pixels of a two-level image with a lower brightness level that are located on the edge of the image.

Let us consider, in the most general form, how a histogram of the length of borders can be constructed. During the successive increase in the threshold, the boundary segment of a pair of bordering pixels will be included in the length of the boundaries of the binarized image when passing the lowest brightness value in the considered pair of bordering pixels in a grayscale image. Then, it will be excluded from the length of the boundaries when passing the highest brightness value in the pair under consideration. In the histogram of the border length, these pixel brightness values correspond to two cells, in one of which the pair of pixels in question provides the addition of the length of the segment, and in the other, the subtraction of the length of the segment. Thus, having examined all the boundary segments between adjacent pixels and having increased and decreased their corresponding cells, we can construct a histogram of the length of the boundaries without performing trial binarization.

You can easily see that the number of boundary segments between adjacent pixels in a two-level image to be analyzed is slightly less than twice the number of pixels. To determine the highest and lowest brightness values in a pair of pixels, you need only a few simple arithmetic operations. Moreover, in most modern processors, the operations of calculating the minimum and maximum are atomic processor instructions. All this leads to low computational complexity and speed of building a histogram of the length of the borders. As will be shown below, it can be constructed, for example, by comparing the brightness of each pixel in a section of a banknote with the brightnesses of four neighboring pixels adjacent to it, although there may be other methods with even lower computational costs. Thus, building a histogram of the length of the borders requires four comparison operations per pixel. It is possible that there is an even faster, but still unknown, way to construct a histogram of the length of the borders.

Further data processing in the claimed invention is based on a histogram of the length of the boundaries of the binarization result. The complexity of this processing does not depend on the number of pixels and is determined only by the number of histogram levels. In accordance with the criteria for analyzing the histogram of the border length given below, for each level of the histogram, only a few simple arithmetic operations are necessary. The practically necessary number of histogram levels does not exceed 256 and, as practice shows, can be reduced to 64. Therefore, the computational complexity of the criterion for analyzing the histogram of the border length is lower than the computational complexity of constructing this histogram.

In the prototype, during trial binarization, for each trial threshold value, it is necessary to compare the brightness of each pixel with the value of the trial threshold. In addition, in the prototype for searching for clusters in each test binarized image, it is necessary to compare the brightness value of each pixel with at least four neighboring ones. Thus, in the prototype, for each pixel, the number of comparisons is equal to the number of trial threshold values multiplied by at least 5. In practice, at least 16 trial threshold values are recommended, which is why the prototype requires the number of comparisons per pixel, which is an order of magnitude greater than the number of comparisons required in the claimed invention. To this we must add the cluster analysis procedures, the complexity of which is rather difficult to assess, but which additionally increase the computational complexity of the prototype.

Thus, the computational complexity of the claimed invention is significantly lower than the computational complexity of the prototype. This ensures the achievement of a technical result.

At the same time, the claimed invention provides at least the same accuracy in determining the final binarization threshold as that provided by the prototype. In the most general form, both the claimed invention and the prototype guarantee that the final binarization threshold will be between the minimum and maximum allowable thresholds. In the claimed invention, the criterion for analyzing the histogram of the border length is used for this.

To explain the use of the histogram of the length of the boundaries when finding the final binarization threshold, we consider a phased change in the length of the boundaries of the binarized image as the binarization threshold increases sequentially. At the first stage, there are no black pixels in the binarized image, so the length of their borders is zero. When disparate pixels appear in the second stage, the length of the borders begins to increase sharply. Without loss of generality, we will consider square pixels, and as units of the length of the border we will use the side length of the pixel. Each square pixel has four equal sides, therefore, when it appears in a binarized image, the length of the image borders grows immediately by 4 units. In the third stage, disparate pixels are combined into small clusters. Adding a new pixel to the cluster increases the length of the borders by less than 4 units, however, in general, the growth rate of the length of the borders remains high.

The situation changes in the fourth stage, in which the symbol in the binarized image consists of continuous lines. As the threshold increases, the thickness of the lines increases. However, the length of the boundaries of the lines grows only due to their ends. On extended sections of lines, the length of the boundaries remains practically unchanged. Accordingly, at the fourth stage, the growth of the boundaries of the binarized image is very small. For closed symbols, where the lines have no ends, the growth of borders is practically absent. At the fifth stage, with the manifestation of noise scattered pixels of the background in the binarized image, a sharp increase in the length of the borders begins again at a rate of up to 4 units per pixel.

Summarizing what is described here, we can say that in the second, third and fifth stages, at least part of the cells of the histogram of the border length contains significant positive values associated with the appearance of disparate pixels and the growth of small clusters. In contrast, the average values of the histogram cells of the border length at the fourth stage are very small, since the border growth is small or practically absent. This explicit contrast makes it possible to implement various criteria that find the binarization threshold on an acceptable range of histogram levels from the minimum to the maximum allowable, corresponding to the fourth stage. Any value of the binarization threshold within the allowable range ensures the transmission of the symbol in the binarized image while maintaining its shape. The specific value of the binarization threshold can be found within an acceptable range using additional considerations. The simplest solution may be to select a threshold in the middle of an acceptable interval. The main difference between the prototype and the claimed invention can be described as follows. In the prototype, multiple trial binarization with different threshold values is carried out, its results are analyzed and the final binarization threshold is selected based on the results of the analysis. In the analysis, the geometric features of the clusters are extracted and compared to determine the binarization threshold at which the clusters will almost exclusively be binarized images of symbols. In the claimed invention, a histogram of the length of the boundaries is used to analyze the result of binarization, carried out for the same purpose as in the prototype, but without the need for binarization itself. The lack of trial binarization, ultimately, explains the reduced computational complexity and increased speed of the claimed method in comparison with the prototype.

The threshold procedure can be defined in various ways. In the most common definition of a binarization threshold and a threshold procedure, pixels whose brightness is less than the brightness of the binarization threshold are assigned to a symbol, and the remaining pixels are assigned to the background. For the threshold procedure defined in this way, the transition level of the pixel is one more than the pixel brightness value. With a different definition of the binarization threshold and the threshold procedure, when pixels whose brightness is less than or equal to the brightness of the binarization threshold are assigned to a symbol, the transition level of the pixel is equal to the brightness of this pixel. For the claimed method, the difference in the definitions of the binarization threshold and the threshold procedure does not play a significant role. Since the transition level differs from the pixel brightness level by a constant value, the ratios “greater”, “less” and “equal” for the brightness level between adjacent pixels coincide with the ratios “greater”, “less” and “equal” for the transition level of neighboring pixels. The concept of a transitional level allows a single way to consider various implementations of a threshold procedure.

To construct a histogram of the border length in accordance with the general approach given above, initially assign all the histogram cells of the border length to zero values, and then analyze the boundaries between the bordering pixels, for which, for each pair of bordering pixels of a grayscale image separated by a border line, the brightness of the pixels is compared and determine a less bright and brighter pixel in the pair, and then increase the value of the cell corresponding to the transition level of the less bright pixel in the pair, by the length boundary segment, and reduce the value of the cell corresponding to the transition level of the brighter pixel in the pair by the length of the boundary segment, where the transition level of the pixel is such a level from an ordered set of increasing brightness values at which the value of the corresponding pixel in the two-level image formed by the threshold procedure, changes when the threshold used in the threshold procedure changes to a transitional level from the previous level in the said ordered set, and, for taking into account the additional component of the border length, for each pixel located on the edge of the image, the value of the histogram cell of the border length corresponding to the transition level of this pixel is increased by the length of the pixel side located on the edge of the image.

For simplicity of reasoning, we will talk about black pixels of the symbol and white pixels of the background in the binarized image. The length of the borders in the binarized image is determined by the total length of the boundary segments between adjacent bordering pixels, one of which is black and the other is white.

To form a histogram of the length of the boundaries, separately considered are all pairs of adjacent pixels of a grayscale image. As the binarization threshold increases, the result of the threshold procedure for a pixel in a pair changes at the moment when the binarization threshold reaches the transition level of this pixel. Such a change, referred to here as a transition, contributes to the value of the histogram cell of the border length, which corresponds to the transition level. When the transition causes the pixels in the two-level binarized image corresponding to the pair under consideration to receive different values, the length of the borders in the binarized image increases by the length of the boundary segment. To reflect this in the histogram of the length of the borders, the length of the boundary segment should be added to the cell corresponding to the transitional level of the changed pixel. On the contrary, if, as a result of the transition, the pixel values in the two-level binarized image become the same, then the length of the boundaries decreases and the length of the boundary segment must be subtracted from the cell corresponding to the transition level of the changed pixel. At the same transition levels, as a result of a synchronous transition of pixels, the length of the borders does not change, and there is no need to modify the histogram cells of the border length.

In order to implement the described logic of modifying the value of the cells of the histogram of the border length, the brightness of the pixels composing a pair of adjacent pixels is compared. Based on the comparison results, when the brightness levels of the bordering pixels are different, then the value of two histogram cells corresponding to the transitional pixel levels is changed by the length of the boundary segment using the corresponding sign. With the equality of the brightness levels of pixels and, accordingly, their transition levels, do not change a single cell of the histogram.

In order to take into account the additional component of the border length, for each pixel located on the edge of the image, the value of the histogram cell of the border length corresponding to the transition level of this pixel is increased by the length of the pixel side located on the edge of the image. Such an implementation of accounting is based on the assumption that outside the region in question there are pixels of obviously higher brightness than any pixel in the region of the image containing the symbol. Due to this, the border of the dark symbol, located on the border of the site, is considered in exactly the same way as the border of this symbol with a brighter background. Namely, the side of the pixel located on the border of the considered area, as the threshold grows, is included in the total length of the borders, as soon as the specified pixel, with binarization, goes from a high level of brightness to low.

To speed up the formation of a histogram of the length of the borders, a comparison of the brightness levels of the pixels and the corresponding modification of the histogram of the length of the borders can be grouped in the most efficient way. In one implementation of the method, first set the cumulative amount to zero, and then if there is a bordering pixel located in the line with a lower number, and the brightness of the pixel in question is greater than or equal to the brightness of the bordering pixel, then the cumulative amount is increased by the length of the boundary segment, if any bordering pixel with a lower number, located on the same line, and the brightness of the pixel in question is greater than or equal to the brightness of the bordering pixel, then the cumulative amount is increased by the length of the border segment, if there is a bordering pixel located in the line with a large number, and the brightness of the considered pixel is greater than the brightness of the bordering pixel, then the cumulative sum is increased by the length of the boundary segment, if there is a bordering pixel with a large number located on the same line, and the brightness of the considered the pixel is greater than the brightness of the bordering pixel, then the cumulative amount is increased by the length of the boundary segment, after which the cell of the histogram of the increment of the border length corresponding to the transition an equal considered pixel, by the number 4, from which is subtracted twice previously accumulated amount.

An examination of each pixel is an analysis of the influence of its boundaries with other adjacent pixels on the value in the histogram cell of the increment of boundaries corresponding to the transitional level of the considered pixel. The cumulative sum M corresponds to the number of boundary segments of the considered pixel with less bright bordering pixels. When the binarization threshold increases, the transition of less bright bordering pixels occurs first, and only then does the transition of the considered pixel occur. Therefore, the transition of the considered pixel reduces the length of the boundaries of the binarized image in accordance with the value of the cumulative sum. In the case of brighter bordering pixels, the transition of the considered pixel occurs before the transition of brighter bordering pixels, and increases the length of the borders. The total number of bordering pixels, regardless of their brightness, is 4. Therefore, the negative contribution of the border segment with a darker pixel not only directly reduces the value of the cell of the histogram of the border increment corresponding to the transition level of the pixel in question, but also reduces the number of brighter bordering pixels, increasing the value of the cell. Thus, the negative contribution of the cumulative sum M to the histogram cell of the boundary length must be taken into account twice. As a result, a value of 4-2M characterizes the full contribution of the pixel in question to the specified cell.

For pixels located on the edge of the plot, the absence of a bordering pixel does not increase the cumulative sum M and, due to this, is always taken into account as a boundary segment with a brighter pixel beyond the boundary of the plot.

In two of the four conditions used to change the cumulative sum M, the condition "greater than or equal to" is used for brightness, and the condition "greater than" is used in the remaining two. Due to this, the case of equality of brightness of two adjacent pixels is considered in such a way that the length of the boundary segment between the two specified pixels is taken into account twice and leads, once, to an increase in the corresponding histogram cell of the border length, and another time to its reduction. Due to this, the border between two pixels of equal brightness is not taken into account in the histogram of the border length, since at any binarization threshold it does not enter the border length of the binarized image.

The grouping of actions described here does not change the previously described general approach based on the choice of the highest and lowest brightness level in a pair of adjacent pixels. The order of actions is changing, but not their result. However, grouping actions reduces the total number of arithmetic operations and, without additional operations, takes into account the contribution of pixels located at the boundaries of the plot.

In order to find the final binarization threshold, when using which the binarized image contains continuous lines, the histogram of the border length is analyzed to find the interval of immediately following luminance values, such that the values of the histogram cells corresponding to the brightness values at the indicated interval do not exceed in advance a predetermined limit, and determine the final binarization threshold based on the values of the endpoints of the specified interval in accordance with a predetermined right silt.

The specified method allows you to set the final threshold between the minimum acceptable and maximum acceptable values, based on the previously described fact that the growth of the boundaries of the binarized image in the fourth stage is very small.

However, a significant additional advantage is provided by the preferred implementation of the method in which a luminance histogram of a grayscale image is additionally generated, where the cells correspond to the luminance values from the ordered set used to construct the histogram of the border length, and, for each luminance value in the ordered set, the corresponding cell value is set as the number of pixels of a grayscale image, the transition level of which is equal to the named value of brightness, and to estimate the average thickness the line widths at the test threshold value use a thickness parameter equal to the ratio of the sum of all cells of the histogram of the border length in the interval from the minimum brightness level in the set to the test threshold to the sum of all cells of the brightness histogram in the interval from the minimum brightness level in the set to the test threshold, and, by to a given rule, as a final binarization threshold, a trial threshold is chosen at which the thickness parameter is close to the target value, and the specified value is pre-selected to reduce the reproducibility error Institution symbol in the final result of binarization based on known tracings those characters which may be located within the area of the banknote.

For each value of the test threshold, the sum of all cells of the histogram of the length of the boundaries in the interval from the minimum brightness level in the set to the value of the test threshold is the area of the dark parts of the binary image obtained as a result of binarization with a test threshold. The area is expressed by the number of pixels. The thickness parameter is calculated as the ratio of the sum of all cells of the histogram of the length of the borders in the interval from the minimum brightness level in the set to the test threshold to the sum of all the cells of the histogram of brightness in the interval from the minimum brightness level in the set to the test threshold. That is, the thickness parameter is the ratio of the length of the boundaries of the binarized image to the area of its dark parts.

Symbols consist of lines, which are characterized by a significant excess of length over width. First of all, we consider a separate long straight line of constant thickness located in the area subjected to binarization. In this consideration, for simplicity, initially we will not take into account the discrete pixel structure of the image. The specified straight line, in fact, is a rectangle, one side of which corresponds to the length of the line, and the other corresponds to its width. For such a line, from obvious geometric considerations, the area is equal to the product of length and thickness.

The length of the borders of such a line is the perimeter of the corresponding rectangle. This perimeter is almost exactly equal to twice the line length, since the line width is small in comparison with the length, and the contribution of the sides of the rectangle corresponding to the width can be neglected. Thus, the line length is almost equal to half the length of the line borders. Accordingly, the area of the line S is almost equal to half the product of the length of the boundaries B and the width of the line W.

If we additionally take into account the pixel structure of the image, then each of the sides of the rectangle must be replaced by a broken line consisting of horizontal and vertical segments corresponding to the sides of the pixels. Therefore, the length of the borders is the sum of the lengths of the horizontal and vertical segments in the composition of the broken lines corresponding to the sides of the rectangle. Because of this, the length of the line borders, taking into account the discrete pixel structure of the image, differs from the ideal geometric length of the borders, considered without taking into account the discreteness, by a discreteness coefficient of approximately 1.2 ± 17% and depending on the angle of the line. The minimum value of the discreteness coefficient is obtained in the case of vertical and horizontal orientation of the line, and the maximum is achieved with an inclination of 45 °. These considerations are directly supported by the Pythagorean theorem. When inclined at 45 °, the length of the border, taking into account the discreteness, is formed by the sum of the same vertical and horizontal legs of rectangular triangles formed by the sides of the pixels. This sum is approximately 1.41 times greater than the length of the line boundary without discretization, calculated as the sum of the hypotenuses of these triangles. For any other angles, the difference between the sum of the legs and the hypotenuse is less than named, and is completely absent at angles of 0 ° and 90 °.

. Указанное соотношение, с небольшой потерей точности, сохраняется и для изогнутых линий. Пренебрегая малым вкладом в границы и площадь в местах пересечения линий, это соотношение можно далее распространить на единичный символ либо несколько символов, выполненных одним шрифтом, расположенных в участке изображения, подлежащем бинаризации. Точное значение коэффициента обратной пропорциональности будет варьироваться для различных символов либо их наборов, в зависимости от конкретной формы и ориентации линий в символе либо символах.Thus, taking into account the discreteness, the area of the line S is almost equal to (0.6 ± 17%) BW. Due to this, the thickness parameter P, defined as the ratio of the length of the borders to the area, is inversely proportional to the line thickness:

. The indicated ratio, with a slight loss of accuracy, is also preserved for curved lines. Neglecting the small contribution to the boundaries and the area at the intersection of lines, this ratio can be further extended to a single character or several characters made in one font, located in the area of the image to be binarized. The exact value of the coefficient of inverse proportionality will vary for different characters or their sets, depending on the specific shape and orientation of the lines in the character or characters.

There are complex type fonts in which the line thickness can vary from one end of the line to the other. For such fonts, the thickness parameter is inversely proportional to the average line thickness. In general, the thickness parameter corresponds to an artistic characteristic called font saturation, which changes on a light to bold scale. A lighter font has a smaller line thickness and, accordingly, a larger symbol thickness parameter, while a bold font has a larger line thickness and a smaller symbol thickness parameter. When composing the styles of various characters of the same font, the designer tries to ensure the constancy of saturation for all characters, as this provides a harmonious perception of the text typed in the font. This explains the close values of the thickness parameter for characters that make up one font.

In a preferred embodiment, according to a given rule, a trial threshold is selected as the final binarization threshold at which the thickness parameter is close to a predetermined target value. The specified target value is chosen to reduce the error in the reproduction of the symbol in the final result of binarization, and this choice is made on the basis of the known styles of those symbols that can be located within the banknote section. The target value of the thickness parameter, which provides a reduction in the error of reproduction, can be set by theoretical analysis of reference images of characters of a known font or trial binarization of images of characters of this font.

Theoretical analysis allows you to calculate the area and length of the borders for a known reference binary image of each character in the font. From the calculated values, a generalized value of the thickness parameter should be found, which, in general, corresponds to the characters of the font. Using such a generalized thickness parameter as the target, in turn, will ensure that the final result of binarization is close to the corresponding reference image.

With trial binarization, you should get several scanned images of each of the characters in a known font! Next, each scanned image must be binarized many times using different binarization thresholds, and, using the obtained binary images, find the corresponding thickness parameter values. From various values of the thickness parameter, it is necessary to select and average those that provide the most accurate reproduction of characters in the binary image. The average value of the thickness parameter should be used as the target.

The preferred implementation allows you to find the final threshold, close to the best. With such a final threshold, in a binarized image, with an accuracy of approximately ± 17%, the average thickness of the lines characteristic of the known font with which the character or characters is made will be achieved. By preserving the characteristic thickness of the lines, the similarity between the final binarization results of the same symbol located on different banknotes is increased, and the effect of differences in pollution during banknote wear is reduced. The increase in similarity, expressed, in particular, in reducing the spread of distances between the characteristic points of the symbol, allows recognition with a reduced level of errors. It is important to note that for the final result of binarizing images of different instances of a symbol of the same style, the difference in the average line thickness between the instances is significantly less than ± 17%. This happens because the discreteness coefficient depends on the shape of the lines and changes little with the same line shape in the symbol. Therefore, although the distortion of the thickness of the lines of a certain symbol, in comparison with symbols of other styles, will be within approximately ± 17%, but it will vary little from one instance of the symbol to another instance of the symbol of the same style. In other words, the preferred method provides increased repeatability of the average line thickness in the final binarization result for symbols of the same style. It should be borne in mind that increased repeatability cannot be achieved when more than one character is located in the image area, since the resolution coefficient will vary depending on the character set.

In FIG. 1 shows a grayscale image of the serial number of the banknote to be binarized (A) and the results of its binarization using various threshold values (B - H).

In FIG. Figure 2 shows a histogram of the length of the borders (A) and a histogram of the brightness (B) of the grayscale image of the serial number. In the same place, as a function of the applied binarization threshold, the dependences of the length of the boundaries and thickness parameter (A), as well as the area of black pixels (B) of the binarized image are given.

In FIG. Figure 3 shows the contribution of the boundary between two pixels to the total length of the boundaries at different values of the binarization threshold, and the contribution of these pixels to the histogram of the length of the boundaries.

In FIG. 4 is a flow chart for calculating a histogram of border lengths.

In FIG. 5 shows various options for the formation of the contribution of a pixel to the total length of the borders at various brightness ratios of this pixel with respect to the brightness of the adjacent pixels.

In FIG. 6 shows a block diagram of the application of the claimed method in a device for processing banknotes.

In an example implementation of the claimed method, binarization of the image of the serial number of the banknote is carried out. To illustrate, a grayscale image of a serial number of a banknote of 200 Israeli shekels, shown in FIG. 1A. This image contains a number containing 10 digits, as well as a background in the vicinity of the number. Each image pixel is represented by a brightness value in the range from 0 (darkest) to 255 (lightest).

A luminance histogram 3 of this image is shown in FIG. 2 and is three-mode. The background image has a high optical density, which varies greatly in different parts of the image. The two right modes on the histogram 3 brightness relate to the background and are separated by a deep dip, which makes it difficult to choose the threshold based solely on the histogram. When using the Father’s algorithm, the separating point on such banknotes is often taken not as a section between the background and the symbol, but as a gap separating the two background modes. This often leads to incorrect binarization and the inability to recognize the serial number.

In binarization using a fixed threshold L, all pixels whose brightness is less than or equal to L. are marked black in the resulting binarized image. In the threshold function selected in this way, the transition level of the pixel is equal to its brightness. By using different threshold values L, binarized images shown in FIG. 1B - FIG. 1H, in order of increasing threshold. The values of the threshold L used are shown to the right of each binarized image.

For the binarized image in FIG. 1B, the binarization threshold L = 20 is higher than the darkest pixels of the symbol, which is why scattered clusters of black pixels related to the symbol begin to appear in the binarized image. Such a threshold corresponds to the second stage described previously. The image in FIG. 1C (L = 27) corresponds to the moment when line breaks disappear almost completely, with the exception of single drops of black pixels that do not affect the quality of recognition. This image corresponds to the beginning of the fourth stage and the minimum permissible binarization threshold. The binarized images obtained in the fourth stage allow for confident character recognition. Binarized images obtained in the previous steps are not suitable for recognition due to massive violation of the integrity of the lines of the symbol.

As mentioned earlier, during the fourth stage, a further increase in the threshold leads to the fact that the pixels in the uncertainty band gradually change from the background to the symbol, which is accompanied by an increase in the thickness of the lines while maintaining the structure of the symbol. The images in FIG. 1D corresponds to the best binarization threshold L = 32, at which the average thickness of the symbol lines in the binarized image is equal to the average thickness of the lines of the original symbol. With a further increase in the threshold, the increase in the thickness of the lines continues while maintaining the outline, as shown in FIG. 1E. This image corresponds to the maximum permissible threshold L = 37 and the completion of stage 4. At the maximum allowable threshold, the manifestation of individual single black pixels at the location of the background is possible, which does not lead to difficulty in character recognition. In a fifth step, as shown in FIG. 1F (L = 45), when the threshold is increased above the maximum allowed, some background pixels in the digital image appear below the threshold and appear in the binarized image as black pixels. Due to this, noise clusters of black pixels appear in the spaces. In a sixth step, as shown in FIG. 1G, with a threshold value of L = 64, which is much larger than the maximum allowable, noise clusters merge with the lines of the symbol. In a seventh step to which the image in FIG. 1H, noise clusters and symbol lines completely merge with each other, forming a solid black filling of the binarized image with white inclusions. Starting from the fifth stage, the binarized image is poorly suited for character recognition due to the mass merging of the elements of the symbol with areas of the background. Thus, the acceptable threshold interval L is from 27 to 37, and the best threshold value L = 32 lies approximately in the middle of this interval. To determine the smallest allowable, largest allowable, and best threshold value, a histogram of the border length is constructed based on the grayscale image and then it is analyzed. In FIG. 2A shows a border length histogram 1 constructed based on the grayscale image shown in FIG. 1A.

Obtaining a histogram 1 of the boundary length is based on the observation, which is illustrated in FIG. 3. Consider any pair of pixels of a grayscale image having a common border. As an example, in FIG. 3, these pixels have brightnesses of 10 and 20 and are separated by a horizontal boundary line between them. For 0 <L <10, as a result of binarization, both pixels in the binarized image are displayed as white, the border between black and white along the pixel boundary in the pair does not pass, and the contribution of this pair to the total border length d = 0. For 10≤L <20, a pixel with a brightness of 10 is displayed as black, and a pixel with a brightness of 20 is displayed as white, so that the border between black and white passes along the border line between the pixels, and the contribution of this pair of pixels to the total length of the borders is d = l. At 20≤L≤255, both pixels in the binarized image are displayed as black, and the contribution of this pair to the total length of the boundaries is d = 0. That is, with a successive increase in L, the border between black and white in the binarized image begins to pass along the border line between pixels in a pair at L = 10, and ceases to pass along the border line between pixels in a pair at L = 20. From this we can conclude that in order to reflect the change in the contribution of the boundary segment between pixels to the histogram of the length of the boundaries, it is necessary to increase the cell of this histogram HB (10) by 1 and reduce the cell HB (20) by 1.

More generally, a pair of bordering pixels contributes +1 (increment) to the histogram cell corresponding to the lowest pixel brightness in the pair, and -1 (decrement) to the histogram cell corresponding to the highest pixel brightness in the pair. To build a histogram of the length of the borders of a grayscale image, you need to consider all pairs of bordering pixels in this way, and additionally take into account the increase in the length of the borders at the edges of the image. To take into account the contribution of all possible pairs, we can consider the boundaries of each pixel not with all neighbors, but only with the neighbors on the right and bottom, which eliminates the erroneous consideration of each border twice. For most pixels in the image, with the exception of those emerging on the right and bottom edges of the image, you need to evaluate their contribution to two pairs with adjacent pixels. Thus, the number of pairs to be considered is almost double the number of pixels in a grayscale image. For pixels at the edge of the image, the positive contribution of the side of the pixel extending along the edge of the image to the total length of the borders occurs at the moment of transition, that is, when the binarization threshold becomes equal to the brightness of such a pixel. Accordingly, for each pixel with brightness I located on the edge of the image, it is necessary to additionally increase the cell HB (I) by the number of sides of the pixel facing the edge of the image. For each pair of bordering pixels, using the observation described above, you need to make two calls to the histogram of the border length. For almost all pixels of the image, therefore, you need 4 accesses to the histogram of the border length. It is possible, having retained the general principle of taking into account the contribution of each pair of bordering pixels to the histogram of the border length, in order to improve performance, reduce the number of calls to the histogram of the border length. Such an opportunity was used in an example implementation of the claimed method and will be described below.

To obtain a histogram 1 of the boundary length, a computational process is used, the block diagram of which is shown in FIG. 4. In this process, all pixels of a grayscale image are individually considered. When processing images, it is generally accepted that the row with the lower number is above the row in question and the pixel with the lower number to the left of the pixel in question. This numbering order is used in this description. For each pixel, first, the cumulative sum M is reset to zero (step 102), and then the brightness ratios I of the pixel in question and the four pixels adjacent to it are analyzed from above, left, bottom, and right (steps 103 - 106). The cumulative sum M is incremented each time when the brightness I of the pixel in question is greater than or equal to the brightness of the bordering pixel from above or to the left. Similarly, the cumulative sum M is incremented each time when the brightness I of the pixel in question is greater than the brightness of the bordering pixel from below or to the right.

Let us examine in more detail what happens at the transition moment, when the binarization threshold L changes from the value (I-1) to the transition value I, that is, when the pixel in question changes color from white to black in the binarized image. We will denote by D the increase in the length of the boundaries of the binarized image extending along all four sides of the considered pixel at the moment of its transition from the white state to the black one. D is expressed in units equal to the length of the side of the pixel. In FIG. 5, using the values indicated in the respective squares of the pixels, the brightness of the adjacent pixels is shown. In FIG. 4A, all 4 pixels are brighter than the considered one, that is, M = 0. Upon transition, all 4 sides of the pixel in question increase the length of the borders by D = 4.

In the case of FIG. 5B, out of 4 adjacent pixels, one is darker than the considered one and three are lighter, that is, M = 1. Before the transition, the border shown by the dashed line passes between the darker pixel and the considered pixel, and it disappears at the moment of transition. At the same time, between the three lighter pixels and the pixel in question, at the moment of transition, three boundary segments are shown, shown by the bold line. In total, this increases the length of the boundaries by D = 2.

In the case of FIG. 5C, out of 4 adjacent pixels, two are darker than the considered one and two are lighter, that is, M = 2. During the transition, two boundary segments disappear and two appear, which gives the final value D = 0. And finally, in FIG. 5D and FIG. 5E shows cases with M = 3 and M = 4, respectively. Upon transition, more boundaries disappear than are added, which gives D = -2 and D = -4, respectively.

In general, for all the cases described, the formula D = 4-2M shown in FIG. 5F. The calculation by this formula is performed at step 107, and the obtained value D is added to the cell HB (I) of the histogram 1 of the border length corresponding to the brightness value I of the pixel in question.

The neighborhood of the considered pixel with a pixel with the same brightness I requires special consideration. At no threshold of binarization can a boundary be formed between them, since they make a transition at the same time. Based on the observation shown in FIG. 3, this pair of pixels simultaneously gives both the increment and decrement of the HB (I) cell, and therefore does not contribute to the histogram of the border length at all. To correctly handle this case, at steps 103 - 106, the comparison "greater than or equal to" is used for the bordering pixel on the top or left, and the comparison "greater" is used for the bordering pixel on the bottom or right. One of the pair of adjacent pixels appears to the left of the other or to the bottom of the other. Both pixels, in step 107, change the cell HB (I) of the histogram 1 of the border length. Due to the difference in the comparison used to arrange a pair of pixels from top to bottom or from left to right, the contributions of these pixels to the HB (I) cell turn out to be opposite. One of the pixels gives a contribution of the length of the boundary segment between them, equal to 1, and the other pixel gives a contribution of -1. As a result of summing these contributions, the border between pixels of equal brightness I does not at all affect the cell HB (I) of histogram 1 of the border length. Also, you need to consider the influence of those sides of the pixel in question that go to the edge of the image. If the side of the considered pixel goes to the edge of the image, the corresponding bordering pixel is absent, and in the corresponding step 103-106 there is no increment of the value of M. Accordingly, at the moment of the transition of the considered pixel, its side facing the edge of the image is always added to the HB cell ( I) histograms of 1 border length. This corresponds to the formula D = 4-2M, calculated in step 107. The described effect is equivalent to the fact that as if beyond the edge of the image there were pixels having a brightness greater than any possible pixel brightness in the image itself. If the symbol line is close to the edge of the image, its border, including the portion along the edge, is fully taken into account in the total length of the boundaries of the binarized image.

Presented in FIG. 4, the process of constructing a histogram of the length of the boundaries is optimized for speed, but performs the same actions on the histogram of the length of the borders, which are depicted in FIG. 3. The process of FIG. 4 groups the actions associated with the pairs where each pixel enters, so as to reduce the total number of calls to the histogram cells of the border length stored in the processor RAM. When considering each pixel, the histogram cell is accessed only once, and decrement operations are not used. Increment operations are performed for the same variable M, which can be stored in the processor register and require minimal time increment. In addition, the process of FIG. 4 allows you to calculate an additional component to take into account the length of those sides of the pixels that are on the edge of the image, without applying additional steps.

Consider the obtained histogram 1 of the length of the boundaries. Using line 2 shows a graph of the dependence of the length of the boundaries B (L). The length of the boundaries B (L) for a given level of brightness L is the sum of all cells of the histogram 1 of the length of the boundaries from zero to a given level L.

For convenience of consideration, the brightness range of the possible values of the binarization threshold L is divided into individual intervals using dashed lines 6-13. For brightness values to the left of line 6, there are completely no pixels with such values that can be seen from the zero values of the histogram 3 of brightness and histogram 1 of the length of the borders. Between lines 6 and 7 there are brightness values corresponding to the second and third stages, where the symbol appears in the binarized image only partially. In particular, there is a binarization threshold for image acquisition in FIG. 1B.

The minimum permissible binarization level L = 27 and the image in FIG. 1C corresponds to line 7. The maximum allowable binarization level L = 37 and the image in FIG. 1E corresponds to line 9. The best binarization level is L = 32 and the image in FIG. 1D corresponds to line 8.

As already described earlier, at the fourth stage of increasing the binarization threshold, in the interval between lines 7 and 9, the length of the boundaries of the binarized image B (L) grows to a minimum with increasing L. Accordingly, the cells HB (L) of histogram 1 of the length of the boundaries in this section take very small values close to zero. Therefore, to find the maximum and minimum acceptable levels, it is enough to find the interval where the absolute value of the cells HB (L) of histogram 1 of the boundary length will be less than a certain empirically specified limit. In our example, such a limit of HB _lim is 5.

In the interval between lines 9 and 10, corresponding to the fifth stage, there is a sharp increase in the values of the histogram cells 1 of the boundary length due to the appearance of noise clusters in the background image area. To the right of line 10, the HB (L) cells become negative, which corresponds to a decrease in the length of the B (L) boundaries due to the fusion of black clusters with each other. Fluctuations in the values of cells between lines 10 and 11, and 11 and 12 are associated with the two-mode nature of the background image. Line 13 separates the portion of the luminance interval to the right of which all the pixels in the binarized image are black and the length of the borders B (L) is constant and equal to the length of the edge of the image. This interval corresponds to the eighth stage.

In a simple embodiment of the claimed invention, the minimum and maximum acceptable levels of binarization corresponding to lines 7 and 9 are found. For this, it is not necessary to consider the behavior of the histogram of the length of the boundaries in the section between lines 9 and 13 corresponding to steps five through eight. Only the interval corresponding to the fourth stage should be sought. To do this, they begin to sort out the values of brightness levels from line 6 in the direction of increase. At the same time, they are looking for the interval at which the absolute value of the cells | HB (L) | histograms 1 of the length of the boundaries are less than the empirically specified limit HB _lim . As soon as the value of | HB (L) | for several consecutive levels will go above the limit of HB _lim , you need to finish the search. The best level of binarization is found as the middle of the found interval, which corresponds to the arithmetic average of the minimum and maximum acceptable levels of binarization. The found best level is used to binarize a grayscale image (see Fig. 1A) and obtain the final result of binarization (see Fig. 1D).

.However, this option does not rely on the thickness of the lines in the binarized image, and does not allow to realize the additional advantage associated with the fact that the font used to print the serial number is known in advance. A more complex, but also more effective embodiment of the invention uses the concept of the line thickness parameter introduced earlier. To calculate the thickness parameter P (L), the length of the boundaries B (L) and the number of black pixels S (L) are used. The number of black pixels S (L) is shown by curve 4 in FIG. 2B, and for a given level, L is defined as the sum of all cells of the histogram 3 of brightness from level zero to L. By definition, the line thickness parameter is calculated as

.

Graph P (L) is shown by curve 5 in FIG. 2A. It is not defined at a denominator of zero, that is, to the left of line 6. In the area between lines 6 and 7, the graph falls off from the initial value of 4.0. The initial value of 4.0 is determined by the ratio of the length of the borders to the area for isolated black pixels, which first appear in the binarized image as the threshold grows. The decline of the graph corresponds to the formation of clusters of black pixels, the border of which grows slower than the area. From these clusters, when approaching the minimum acceptable threshold value, symbol lines are gradually formed. The decrease in the number of line breaks of the symbol corresponds to the ongoing decline of the graph.

Between lines 7 and 8, in the section of the fourth stage, the symbol is binarized without significant line breaks and manifestations of background pixels. The P (L) graph decreases, reflecting the increase in the thickness of the symbol lines as L. grows. The local maximum of the P (L) graph between lines 9 and 10 occurs due to the appearance of small isolated black background pixels in the fifth stage. These pixels significantly increase the length of the border. With further growth of L, the graph continues to decline and reaches a very small constant value after line 13.

By repeatedly binarizing reference images of various characters of the numerals of the font used to print the characters of the serial number of Israeli shekels, the optimal value of the line width parameter was determined to be P _opt = 0.72. It is noted in FIG. 2A by dashed line 15. With this value, the average line thickness in the binarized image corresponds to the average line thickness of the print font standard.

To find the best binarization threshold, move along the level scale from line 6 in the direction of increasing L, and P (L) is calculated for each level. As soon as P (L) becomes less than or equal to the optimal value P _{opt of} the line width parameter, the search is stopped and the last value of L is used as the best binarization threshold. In FIG. 3A shows a search stop point 14. The found best level is used for binarization of the grayscale image (Fig. 1A) and obtaining the final result of binarization (Fig. 1D).

Both simple and complex implementations, during the search for the binarization threshold, do not reach the threshold values at which black pixels corresponding to the background appear to a considerable degree in the binarized image. This ensures that the binarization result is independent of the background features. If the background is contrasting and / or multi-mode, but, as happens in most cases, it can be separated from the symbol using a global threshold, then both implementation options allow you to find such a threshold.

The results of simple and complex versions differ in processing pixels in the zone of uncertainty around the symbol. A simple option defines the threshold as the arithmetic mean of the minimum and maximum acceptable values. This approximate solution ensures that in the binarized image there are practically no line breaks of the symbol and black noise pixels of the background, but it does not necessarily correspond to the standard of the average thickness of the symbol lines. A complex option, in addition to the absence of line breaks of the symbol and black noise pixels of the background, ensures compliance with the standard of the average thickness of the symbol lines. By maintaining the reference line thickness, character recognition can be performed with higher confidence. Simple and complex implementation options differ slightly in execution time, since most of the time is spent on building a histogram 1 of the border length. The execution time of any of these options is a fraction of the execution time of the method described in the prototype, since binarization using the found threshold is carried out once, and not many times.

The banknote processing sequence using the inventive method in a counting and sorting machine is shown in FIG. 6. The machine provides transportation of banknotes from the supply pocket, while scanning and recognition is carried out with the determination of the type of banknote (step 201). If the banknote is not recognized, then a decision is made (step 210) to send the banknote to the reject pocket. This solution is implemented using an electromechanical redirecting unit located in the banknote path of a counting and sorting machine. If the banknote is recognized, then a digital image of the serial number is extracted from the scanned banknote image and rotated to compensate for skew (step 203). When selecting images, they rely on the known location of the serial number on a particular type of banknote.

Then, a histogram of the length of the boundaries is constructed using the sequence of steps 100 to 108 (step 204). Next, find the best threshold (step 205) using the simple or complex option described earlier. If a complex version is used, at step 205, before performing a threshold search, a luminance histogram is additionally constructed. The best threshold found is used to binarize the image (step 206).

The resulting binarized image is recognized (step 207) to obtain a serial number string. If the serial number is not recognized (step 208), then a decision is made on the direction of the banknote in the reject pocket (step 210). If the symbol is recognized, then send the banknote to the receiving pocket (step 209) and write the serial number from the array-string into the electronic sorting report (step 211). The described processing (steps 201 - 211) is repeated for all banknotes located in the feed pocket of the counting and sorting machine. After that, the recognized banknotes with the successfully recognized serial number are in the receiving pocket of the machine, and the electronic report contains a list of serial numbers of these banknotes. Banknotes that were not recognized, or in the serial number of which not all characters were recognized, are in the pocket of rejection. An electronic report containing a list of serial numbers can then be used to track the movement of banknotes in circulation.

The methods used to recognize the banknote in step 201, extract the image of the serial number and rotate it to compensate for skew in step 203, and recognize the serial number in step 207 are well known to those skilled in the art for processing banknotes and are widely described in the patent literature.

The method disclosed herein is not limited to the use of the entire serial number of a banknote for binarization. It can also be used in devices where the grayscale image of the serial number is first divided into separate characters containing one character each. In this case, the described method of binarization can be applied to the image of each familiarity individually. Namely, for each image of familiarity, a histogram of the length of the boundaries, finding the best threshold, and binarization using the found best threshold should be constructed. The resulting binarized images should then be used to recognize individual characters. Application of the claimed method to individual characters, in many cases, allows you to reliably separate the characters from the background even in cases where high-quality binarization of the entire serial number using the global threshold is impossible. A similar problem is typical for banknotes where the serial number is printed in translucent ink on a dark contrasting background. The brightness of the pixels of various symbols printed with translucent ink strongly depends on the brightness of the background and varies for different familiarities, which leads to an extension of the overall gradation interval of brightness of the pixels of the symbol. As a result, often there is a strong overlap of gradation intervals of brightness of characters and background within the entire number. The claimed method allows to successfully binarize the characters individually, provided that there is no significant overlap of the gradation intervals of the brightness of the symbol and background within each of the familiarities.

Claims (31)

1. A method for binarizing symbol images on a banknote, in which a grayscale image of a section of a banknote containing at least one symbol to be binarized is obtained, in which each pixel is characterized by a brightness level, and a final binarization threshold is determined and a final binarization result is generated, for which a threshold procedure using the final binarization threshold, where the threshold procedure is the process of forming a two-level image in which each pixel corresponds to a grayscale image pixel, and has one of two possible brightness levels, assigned based on a comparison of the brightness of the corresponding pixel grayscale image with a given threshold for binarization, while to determine the final binarization threshold, a histogram of the length of the boundaries of the specified grayscale image is built, why set an ordered set of increasing brightness values possible in a grayscale image, and, for all brightness values in the specified set, starting from the value following the smallest, set the value of the corresponding cell of the histogram of the border length as an increment of the length of the borders of the two-level image obtained as a result of the threshold procedure, corresponding to the increment of the threshold used in the threshold procedure to the considered brightness value from the previous brightness value in the specified set, moreover, the length of the boundaries of a two-level image is calculated as the sum of the lengths of the segments, each of which is a boundary segment of two adjacent pixels with different brightness levels, plus an additional component to take into account the length of those sides of the pixels that are on the edge of the image, and the specified histogram of length is analyzed according to a given criterion boundaries for finding the final binarization threshold, when using which the binarization result contains continuous lines contained in a grayscale image zhenii banknote portion. 2. The method according to claim 1, in which the additional component of the length of the borders of a two-level image is the sum of the lengths of those sides of the pixels of a two-level image with a lower level of brightness that are on the edge of the image. 3. The method according to p. 2, in which to calculate the histogram of the length of the borders of the grayscale image initially assign to all cells of the histogram of the length of the boundaries zero values, and then analyze the boundaries between adjacent pixels, for which, for each pair of adjacent pixels of a grayscale image, separated by a boundary segment, they compare the brightness of the pixels and determine a less bright and brighter pixel in a pair, then increase the value of the cell corresponding to the transition level of the less bright pixel in the pair by the length of the boundary segment, and reduce the value of the cell corresponding to the transition level of the brighter pixel in the pair by the length of the boundary segment, the transitional level of the pixel is such a level from an ordered set of increasing brightness values at which the value of the corresponding pixel in the two-level image formed by the threshold procedure changes when the threshold used in the threshold procedure changes to a transition level from the previous level in the said ordered set , and, to account for the additional component of the border length, for each pixel located on the edge of the image, increase the cell value of the histogram of the border length corresponding to the transition level of this pixel by the length of the side of the pixel located on the edge of the image. 4. The method according to p. 3, in which, for joint analysis of the length of the borders and taking into account an additional component of the length of the borders, consider all the pixels of the grayscale image, so that for each considered pixel determine the cumulative sum corresponding to the number of borders with less bright pixels, first set the cumulative amount to zero, and then if there is a bordering pixel located in the line with a lower number, and the brightness of the pixel in question is greater than or equal to the brightness of the gran current pixel, then the cumulative amount is increased by the length of the boundary segment, if there is a bordering pixel with a lower number located on the same line, and the brightness of the pixel in question is greater than or equal to the brightness of the bordering pixel, then the cumulative amount is increased by the length of the boundary segment, if there is a bordering pixel located in the line with a large number, and the brightness of the pixel in question is greater than the brightness of the bordering pixel, then the cumulative amount is increased by the length of the boundary segment, if there is a bordering pixel with a large number located on the same line, and the brightness of the considered pixel is greater than the brightness of the bordering pixel, then the cumulative sum is increased by the length of the boundary segment, after which the cell of the histogram of the increment of borders is changed, corresponding to the transitional level of the considered pixel, by the number 4, from which doubled cumulative amount is previously subtracted. 5. The method according to any one of paragraphs. 1-4, in which, to find the final threshold of binarization, when using which the binarized image contains continuous lines, analyze the histogram of the length of the boundaries for the determination of the interval immediately following each other brightness values, such that the values of the histogram cells corresponding to the brightness values on the specified interval, modulo do not exceed a predetermined limit,  and determine the final binarization threshold based on the values of the endpoints of the specified interval in accordance with a predetermined rule. 6. The method according to any one of paragraphs. 1-4, in which the criterion for finding the final binarization threshold provides a quasi-optimal approximation of the average thickness of the lines in the final binarized image to a given target value. 7. The method according to p. 6, in which additionally form a histogram of the brightness of the grayscale image, where the cells correspond to the brightness values from an ordered set of increasing brightness values used in the construction of a histogram of the length of the borders, and, for each brightness value in the ordered set, the value of the corresponding cell is set as the number of pixels of a grayscale image whose transition level is equal to the named brightness value, and to estimate the average line thickness at the trial threshold value, a thickness parameter equal to the ratio of the sum of all cells of the length histogram is used boundaries in the interval from the minimum brightness level in the set to the test threshold to the sum of all cells of the brightness histogram in the interval from the minimum brightness level in the ordered abortion to a trial threshold, and, according to a given rule, such a trial threshold is selected as the final binarization threshold, at which the thickness parameter is close to the target value, moreover, the specified target value is pre-selected to reduce the error in the reproduction of the character in the final binarization result based on the known styles of those characters that can be located within the banknote section.

Images