KR101140699B1 - Forgery Detection System for Security Printing Resource Based on Digital Forensic Technology and Method Therefor - Google Patents

Abstract

According to the present invention, it is possible to develop a security printed material forgery identification system and method capable of extracting various feature points.
In addition, security printed counterfeit identification system and method using digital forensic technology that can generate a learning model for counterfeit identification by conducting machine learning for each feature point extracted from digital image for a large number of originals and digital image for counterfeit. Can be implemented.
Security imitation forgery identification system according to an embodiment of the present invention is an image conversion unit 10; Feature point extracting unit 20; Machine learning unit 30; And forgery identification unit 40; includes.

Description

Forgery Detection System for Security Printing Resource Based on Digital Forensic Technology and Method Therefor}

The present invention relates to a security imprint forgery identification system and method, and more particularly, to a system and method for identifying forgery of a security print using digital forensic technology.

With the advent of 50,000 won bills and the circulation of various gift certificates, the counterfeiting of security prints is rapidly increasing. The rapid development of scanners and color printers is also accelerating this forgery. Thus, the demand for a system for identifying counterfeit in a general business dealing with cash has surged.

For the identification of counterfeits by scanners and printers, banding defects occur periodically in printed documents due to quasi-periodic fluctuations caused by the angular velocity becoming inconsistent due to the reverse rotation of the gears that rotate the OPC drum. A technique for identifying monochrome laser printers using facts has been introduced. This technique is easy to apply to halftone image prints and achieves good results, but it is difficult to apply to text prints such as banknotes, so new features are needed to identify printers using text prints.

In addition, the previously applied patent 'Kwon Heterogeneous Classification and Counterfeit Identification Counter using CCD Image Sensor (Application No .: 1020030023962)' identifies a counterfeit using a binarization method, but it is difficult to extract various feature points to improve the identification rate. It is difficult to apply to other digital image acquisition devices.

The present invention has an object to solve the above technical problem, and the object of the present invention is to develop a security printed material forgery identification system and method capable of extracting various feature points.

In addition, the security imprint forgery identification system and method of the present invention is to enable the machine learning by the feature points extracted from the digital image for a large number of originals and the digital image for the counterfeit to generate a learning model for identification of counterfeit or not The purpose is.

Security imitation forgery identification system according to an embodiment of the present invention, the image conversion unit for generating a digital image by converting the printed matter into a digital image; A feature point extraction unit for extracting feature points of the digital image; A machine learning unit for performing machine learning using the feature points; And a forgery identification unit for identifying whether the printed article is forged from the machine learning result.

In addition, the image conversion unit, the first image conversion unit for generating the original digital image and the counterfeit digital image by converting the original print and the counterfeit print out of the prints into a digital image; And a second image converting unit converting the test print out of the printed matter into a digital image to generate a test print digital image.

The feature point extracting unit may include: a first feature point extracting unit extracting feature points of the original digital image and the counterfeit original digital image; And a second feature point extractor for extracting feature points of the test print digital image.

One preferred embodiment of the machine learning unit, a feature point selector for selecting some feature points by the SVM (Support Vector Machine) method using a plurality of feature points from the digital image of the original print and the counterfeit print; A learning model generator for generating a learning model of a feature point vector set based on some feature points selected by the feature point selector; And an SVM classifier for selecting a class to be included in the test print by using the feature points of the test print digital image based on the generated learning model.

In addition, the SVM classifier is characterized by applying a Slack variable and a Kernel function to minimize the error rate of the classification function.

In addition, the kernel function is characterized in that the radial basis function (RBF) kernel (Kernel).

Specifically, the feature point selector selects some feature points using a sequential forward flow selection algorithm.

In detail, the first feature point extractor or the second feature point extractor comprises a feature point vector from the extracted feature points, wherein the feature points are discrete wavelet transform-based feature points; Feature point based on image quality measurement tool; Texture-based feature points; Information theory-based features; And dotnet structure analysis-based feature points.

Specifically, the discrete wavelet transform-based feature points may be characterized by statistical features within a high frequency subband; Feature points due to statistical features between high frequency subbands; And a feature point by a correlation coefficient between color channels of a high frequency subband, wherein the high frequency subband is determined by performing a low pass filter, a high pass filter, and a down sampling process on a horizontal direction and a vertical direction of an image. Among the inverse, it characterized in that it comprises three subbands having a high frequency characteristics in any one or more of the transverse direction and the longitudinal direction.

Characteristic point by the statistical feature in the high frequency subband according to a preferred embodiment is characterized in that it comprises the standard deviation, skewness and kurtosis of each step of the high frequency subband after the discrete wavelet transform.

In addition, the feature point by the statistical feature between the high frequency subbands is characterized by including the mean, variance, kurtosis, and skewness of the error of the approximation value of each subband and the value of the subband after the discrete wavelet transform.

Specifically, the feature point by the correlation coefficient between the color channels is between R and G, between R and B, between G and B, between C and M, and between C and Y determined by performing discrete wavelet transform for each color channel. It is characterized in that the correlation coefficient for the high frequency subbands between, between C and K, between M and Y, between M and K, and between Y and K.

According to a preferred embodiment of the present invention, the image quality measurement tool-based feature points include a pixel difference value and a coefficient difference in a frequency domain in a spatial domain between an arbitrary image input and an image from which noise is removed from an image input by a low frequency filter. It is a feature point extracted based on noise using at least one of a value, a correlation coefficient between two inputs, and a difference value between pixels to which a cognitive visual model is applied.

The low frequency filter may include an averaging filter, a Gaussian filter, a median filter, and a wiener filter.

It is preferable that the texture-based feature points include three statistical characteristics of energy, contrast and similar obtained using Gray Level Co-occurrence Matrices (GLCM).

The texture-based feature may further include an average and a standard deviation of the difference between an arbitrary pixel value and an approximation of a pixel.

A preferred embodiment of the information theory-based feature point is characterized in that it includes an entropy value in each of R, G, and B channels, and mutual information values between R and G, G, B, B, and R channels. .

Specifically, the feature point based on the network structure analysis calculates the linear information existing in the image in the form of distance and angle from the origin through Hough transformation, and calculates the frequency of the histogram constructed using the calculated angle information of the straight line. It is characterized by including.

Security printed material forgery identification system according to another embodiment of the present invention, including a print learning device and a print identification device, the print learning device converts the original print and the counterfeit printed material into a digital image of the original digital image and counterfeit copy A first image converting unit generating a digital image; A first feature point extraction unit for extracting feature points of the original digital image and the counterfeit digital image; A feature point selector for selecting some feature points by a SVM (Support Vector Machine) method using a plurality of feature points of the original digital image and the counterfeit digital image; And a learning model generator for generating a learning model of the feature point vector set based on some feature points selected by the feature point selector.

The print identification device may further include: a second image converting unit converting the test print into a digital image to generate a test print digital image; A second feature point extracting unit extracting feature points of the test print digital image; An SVM classifier for selecting a class to be included in the test print by using feature points of the test print digital image based on the generated learning model; And a forgery identification unit for identifying whether the test print is forged from the class of the SVM classifier.

Security imprint forgery identification method by the security imprint forgery identification system according to an embodiment of the present invention, Original image digitizing step of generating the original digital image by converting the original printout to a digital image; A counterfeit digitizing step of generating a counterfeit digital image by converting the counterfeit printed matter into a digital image by the image converting unit; An original feature point extraction step of extracting a feature point of the original digital image by the feature point extraction unit; A counterfeit feature feature extraction step of extracting a feature point of the counterfeit bone digital image by the feature point extractor; A feature point selecting step of the machine learning unit selecting some feature points from the feature points of the original digital image and the feature points of the fake digital image by a SVM (Support Vector Machine) method; A learning model generation step of generating, by the machine learning unit, a learning model of a feature point vector set based on some feature points selected in the feature point selection step; A test print digitizing step of generating a test print digital image by converting the test print into a digital image by the image conversion unit; A test print feature point extracting step of extracting a feature point from the test point digital image by the feature point extracting unit; An SVM classification step of selecting, by the machine learning unit, a class to include the test print using the feature points extracted in the test print feature point extraction step based on the learning model generated in the learning model generation step; And a forgery identification step of identifying whether the test print is forged by the class selected in the SVM classification step in the forgery identification unit.

According to the security imprint forgery identification system and method of the present invention, it is possible to extract various feature points.

In addition, it is possible to implement a security printed counterfeit identification system and method that can generate a learning model for counterfeit identification by conducting machine learning based on feature images extracted from digital images of a large number of originals and digital images of counterfeit text.

1 is a block diagram of a security printed material forgery identification system according to an embodiment of the present invention.
2 is a block diagram of a security printed material forgery identification system according to another embodiment of the present invention.
Figure 3 shows a block diagram of the machine learning portion of the present invention.
Figure 4 is a block diagram of a security print counterfeit identification system according to another embodiment of the present invention.
5 is a flowchart illustrating a security print counterfeit identification method of the present invention.
FIG. 6 shows a filter structure used to obtain discrete wavelet transform based feature points through an example image.
7 shows the printing noise of a forged document with a color printer.
8 illustrates a process of extracting an image quality measuring tool according to an embodiment of the present invention.
9 shows 17 image quality measurement tools used in the present invention.
10 shows three statistical characteristics using GLCM.
Fig. 11 shows corresponding pixels and their neighbor values in the soft area.
FIG. 12 is a diagram of a gray image output by dot printing using a printer 3 and separated into cyan, magenta, and yellow channels.
13 is a dot structure of a cyan area output to different printers.
14 shows a process of network structure analysis.
15 shows a flowchart of the SFFS algorithm.
16 shows the relationship between the number of feature points selected and the discrimination rate using SFFS.
17 is a schematic diagram of an SVM classifier.
18 shows a conceptual diagram of linear separation using a kernel function.

Hereinafter, with reference to the accompanying drawings will be described in detail for the security printed material forgery identification system according to an embodiment of the present invention.

The following examples of the present invention are intended to embody the present invention, but not to limit or limit the scope of the present invention. It will be understood by those of ordinary skill in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.

First, Figure 1 is a block diagram of a security print counterfeit identification system according to an embodiment of the present invention.

As shown in FIG. 1, the security printed matter forgery identification system according to an embodiment of the present invention includes an image conversion unit 10; Feature point extracting unit 20; Machine learning unit 30; And forgery identification unit 40; includes.

In addition, as shown in Figure 2, the printed material forgery identification system according to another embodiment of the present invention includes a first image conversion unit (50); First feature point extraction unit 60; Machine learning unit 30; A second image converter 70; Second feature point extracting unit 80; And forgery identification unit 40; includes.

The forgery identification system of the printed matter of the present invention will be described in detail with reference to FIG. 1 below.

In the case of printed matter, it is necessary to convert it into a digital image in order to apply the present invention. Therefore, the image conversion unit 10 serves to convert the printed matter into a digital image. Examples of the image converter 10 may include a scanner, a digital camera, and the like.

As can be seen from Figures 1 and 2, the image conversion unit 10 converts the original print and the counterfeit print to a digital image to generate a first digital image and a counterfeit print digital image 50 ; And a second image converter 60 converting the test print into a digital image to generate a test print digital image.

In order to determine whether the printed object is forged from the digital image, it is necessary to extract a feature point that may indicate a difference between the original digital image and the fake digital image. The feature point extractor 20 extracts feature points of the digital image.

As can be seen from Figures 1 and 2, the feature point extraction unit 20 is a first feature point extraction unit 60 for extracting feature points of the original digital image and the fake digital image by the first image conversion unit; And a second feature point extractor 80 extracting a feature point of the test print digital image by the second image converter.

The machine learning unit 30, as shown in Figure 3, a feature point selector 31; Learning model generator 32; And an SVM classifier 33, which performs machine learning using the feature points. Through machine learning, we can create appropriate classification functions, and use them to predict the class corresponding to the new input.

Specifically, the feature point selector 31 selects some feature points by a method based on SVM (Support Vector Machine) using a plurality of feature points from the original digital image and the fake digital image. The learning model generator 32 generates a learning model of the feature point vector set based on some feature points selected by the feature point selector 31, and the SVM classifier 33 generates the test print based on the generated learning model. Using the feature of the digital image serves to select the class to be included in the test print.

The feature point selection technique and the SVM technique used in the machine learning unit 30 will be described later in more detail.

Finally, the counterfeit identification unit 40 identifies whether the test print is forged by the class selected by the SVM classifier 33.

In addition, as shown in FIG. 4, the secure printed matter forgery identification system 100 according to another embodiment of the present invention includes a print learning apparatus 110 and a print identification apparatus 120.

The print learning apparatus 110 may include a first image converting unit 50; First feature point extraction unit 60; A feature point selector 31; And a learning model generator 32.

In addition, the printed matter identification device 120, the second image conversion unit 70; Second feature point extracting unit 80; SVM classifier 33 and the counterfeit identification unit 40 is characterized in that it comprises a.

The security print forgery identification method using the security print forgery identification system 100 of the present invention described above will be described with reference to the flowchart of FIG. 5.

That is, the security printed material forgery identification method of the present invention, the image conversion unit 10 converts the original printout into a digital image to generate an original digital image (S110); Counterfeit digitization step (S115) by the image conversion unit 10 converts the counterfeit printed matter to a digital image to generate a fake digital image; An original feature point extraction step (S120) of the feature point extraction unit 20 to extract a feature point of the original digital image; A counterfeit feature feature extraction step (S125) of extracting feature points of the counterfeit bone digital image by the feature point extractor 20; A feature point selection step of the machine learning unit 30 selecting some feature points from the feature points of the original digital image and the feature points of the fake digital image by SVM method; A learning model generation step (S135) of which the machine learning unit 30 generates a learning model of the feature point vector set by the partial feature points selected in step S130; A test printout digitizing step of converting the test printout into a digital image and generating a test printout digital image by the image conversion unit 10 (S140); A test print feature point extracting step (S145) of the feature point extracting unit 20 to extract a feature point from the test print digital image; An SVM classification step (S150) of which the machine learning unit 30 selects a class to include the test print using the feature points extracted in step S145 based on the learning model generated in step S135; And a forgery identification step (S155) of identifying whether or not the forgery of the test prints by the class selected in the step S150 in the forgery identification unit (40).

A feature point extraction technique in the feature point extractor 20, the first feature point extractor 60, and the second feature point extractor 80 of the present invention will be described in detail below.

Feature Point Extraction Technique

In order to identify the forgery of the test object consisting of digital images, a large number of feature points are extracted and the test object is quantified in the form of a vector. In the present invention, a discrete wavelet transform based feature point, image quality measure based feature point, texture based feature point, information theory based feature point and halftone structure analysis (Halftone) Texture Analysis based feature points were used.

The extraction process of each feature point will be described in detail below.

Discrete Wavelet Transform-Based Features

Through the discrete wavelet transform, a given image is divided into a multi-stage high frequency region and a low frequency region. At this time, the high frequency region represents the noise of the image. Since the noise is slightly different for each color printer or color area, the statistics representing them properly can be regarded as feature points representing an image. Discrete wavelet transform is the same as described below.

First, the basis function Ψ _{j, k} (t) of the discrete wavelet can be expressed by Equation 1.

The arbitrary signal f (t) is represented by the combined form of the basis function Ψ _{j, k} (t) and wavelet coefficients a _{j, k} . That is, it can be expressed as Equation 2.

In multiresolution analysis using wavelets, there are two basic functions, namely, the scale function Ψ _{j, k} (t) and the wavelet function Φ _{j, k} (t). c _j and d _j are the elongation coefficients for each function.

6 illustrates a filter structure using an example image. The filters L (z) and H (z) are elongation coefficients corresponding to the scaling functions? (T) and wavelet function? (T), respectively, where L is a low pass filter and H is a high pass filter. The low pass filter and the high pass filter divide the image into two bands in the horizontal direction, and as a result, the downsampling process takes half the data. Next, each of the two down-sampled bands is divided into two vertical bands, and again, a down-sampling process of taking half the data is performed. Through this process, the signal is expressed in four resolutions, LL, LH, HL, and HH.

In the present invention, after the discrete wavelet transform, statistical characteristics within each high frequency subband, that is, LH, HL, and HH, statistical characteristics between the high frequency subbands, and correlation coefficients between color channels in the high frequency subbands are calculated.

Feature points in high frequency subband

As a characteristic point representing noise in subbands, the standard deviation, kurtosis, and skewness of the high frequency subbands of each step after the discrete wavelet transform, that is, LH, HL, and HH regions, are expressed by Equations 4, 5, and 6, respectively. Obtain

Since the image is composed of three channels (Red, Green, Blue), 27xi feature points can be extracted according to the level i of the discrete wavelet transform. One more thing to consider here is that the printer output consists of four channels of CMYK (Cyan, Magenta, Yellow, Black). Therefore, 36xi feature points can be extracted.

Assuming that only one level wavelet feature point is extracted using a region of interest (ROI) of 64x64 to 128x128, feature points within 63 subbands are extracted.

Feature points between high frequency subbands

After the discrete wavelet transform, the coefficient values of each subband may be expressed in the form of linear interpolation between the coefficient values of the neighboring subbands and the neighbor coefficients. Approximate values of LH, HL, and HH in the i th step may be represented by Equations 7, 8, and 9, respectively.

의 값을 구해보자.

의 계수를

로 나타내면

이다.

와

의 최소자승오류는 수학식 10과 같다.first

Let's get the value of.

Coefficient of

If expressed as

to be.

Wow

The least square error of

를 최소화 시키는 이차 비선형 최적화 문제를 풀면

는 수학식 11과 같다.

Solve a quadratic nonlinear optimization problem that minimizes

Is the same as Equation 11.

를 구했으므로

를 구할 수 있고,

와

사이의 오차는

로 나타낸다. 동일한 방법으로

와

를 구하고 각각에 대하여 평균, 분산, 첨도 및 왜도를 구한다. 이는 부대역 사이의 통계적 특성을 드러내는 특징점이라고 할 수 있다. 부대역 내의 특징점과 마찬가지로, R, G, B, C, M, Y, K 각 채널의 LH, HL, HH 영역에서 부대역 사이의 특징점을 구한다. 상기의 과정을 통해서 총 84개의 부대역 사이의 특징점을 추출할 수 있다.
From above

Since we got

You can get

Wow

The error between

Respectively. In the same way

Wow

And calculate the mean, variance, kurtosis, and skewness for each. This is a feature that reveals the statistical characteristics between subbands. Like the feature points in the subbands, the feature points between the subbands in the LH, HL, and HH regions of the R, G, B, C, M, Y, and K channels are obtained. Through the above process, feature points between a total of 84 subbands can be extracted.

Correlation coefficient between color channels

As already explained, after the discrete wavelet transform, the LH, HL, and HH regions represent the noise of the image. This noise has a unique distribution for each color channel. Therefore, the quantified noise between the color image channels is also an important characteristic of the color characteristics of the printer.

In the present invention, the correlation coefficient is used to quantify the noise between the color channels. First, discrete wavelet transform is performed for each color channel. Next, between R and G, between R and B, between G and B, between C and M, between C and Y, between C and K, between M and Y, and between M and K , And calculate the correlation coefficient between Y and K as the feature point vector. The correlation coefficient is shown in Equation 12.

Through the above process, a correlation coefficient between a total of 27 discrete wavelet transform based color channels can be obtained.

Feature Point of Image Quality Measuring Tool

As can be seen from FIG. 7, a document forged with a color printer contains inherent printing noise that does not exist in the original document. The printing noise is a characteristic caused by the characteristics of the dot printing method itself and cannot be arbitrarily reduced or eliminated. In fact, the printing noise of the counterfeit can be extracted and compared to the original to show distinct values.

와 저주파 필터를 사용하여 잡음을 제거한 이미지

를 입력으로 받아 공간 영역에서 픽셀의 차이 값, 주파수 영역에서 계수의 차이 값, 두 입력의 상관 계수 그리고 인지 시각 모델을 적용한 픽셀의 차이 값으로 분류되는 17가지의 척도를 이용하여 잡음을 측정하였다. 이처럼 다양한 척도를 이용하여 이미지를 정량화한 벡터는 이미지를 구분할 수 있는 좋은 특징점이 된다. However, images cannot be quantified on a single basis. There are several types of image quality measures (IQM) that measure images based on various criteria. In the present invention, a given image

And noise-free images using low-frequency filters

The noise was measured using 17 scales, which are classified into the difference value of the pixel in the spatial domain, the difference value of the coefficient in the frequency domain, the correlation coefficient of the two inputs, and the difference value of the pixel using the cognitive visual model. As such, a vector that quantifies an image using various scales is a good feature to distinguish the images.

에 대하여 잡음을 제거하는 방법을 변경해 가며, 총 4가지의

를 생성하였다. 이후

와 서로 다른

을 이용하여 4 세트의 화질측정도구를 추출하였다. 도 8은 상기의 화질측정도구를 추출하는 과정을 나타낸다.Also in the present invention a given image

We are changing the method of removing noise with respect to

Generated. after

And different

Four sets of image quality measurement tools were extracted. 8 shows a process of extracting the image quality measuring tool.

The noise filters used in the process of extracting the image quality measurement tool are four types: averaging filter (3x3), Gaussian filter (3x3), intermediate value filter (3x3), and Wiener filter (3x3).

The process of extracting the image quality measuring tool is as follows.

에 대하여 상기의 4종류의 필터를 사용하여 화질측정도구를 적용하여 잡음 기반의 특징점을 추출하였다. D1~D3는 픽셀 차이 값을 나타낸다. C1~C5는 상관 계수를, S1~S6은 주파수 영역에서 계수의 차이 값이다. 마지막으로 H1~H3은 인지 시각 모델 U(x)를 적용한 이후 추출한 특징점이다. 4가지 잡음제거 필터를 이용하여 17가지의 화질측정도구를 추출하면 총 68개의 특징점을 구할 수 있다. 도 9에 17가지의 화질측정도구를 나타낸다.
Different filters

The noise-based feature points were extracted by applying the image quality measurement tool using the above four types of filters. D1 to D3 represent pixel difference values. C1 to C5 are correlation coefficients, and S1 to S6 are difference values of coefficients in the frequency domain. Finally, H1 to H3 are feature points extracted after applying the cognitive visual model U (x). A total of 68 feature points can be obtained by extracting 17 image quality measurement tools using four noise reduction filters. Fig. 9 shows 17 image quality measurement tools.

Texture Based Features

The texture can be defined as a function of the change in pixel value in a specific spatial region. Forged documents will have a different texture than the original, as can be seen in Figure 7 due to printing noise. Therefore, the statistical properties of these textures can be another feature point.

In the algorithm used in the present invention, after obtaining GLCM (Gray Level Co-occurrence Matrices), the statistical characteristics of GLCM are extracted and considered as feature points. GLCM Pd of GxG with respect to the direction vector d is defined as in Equation 13. In this case, G represents the number of colors used to represent a black and white image and generally uses 256.

That is, it can be represented by a two-dimensional histogram according to the direction vector d. By observing the constructed GLCM, we can know the characteristics of a given image. For example, if most values of GLCM are centered on a diagonal line, this means that there are many similarities in the image in the direction of the direction vector d. Here, three statistical features shown in FIG. 10 are obtained.

In the present invention, d (0,1), d (1,0), d (1,1), d (1, -1) to d (0,4), d (4,0), d (4, 4) Three statistical properties, 48 feature points, were extracted for 16 direction vectors up to d (4, -4).

는 주어진 계수

와 주변 픽셀 값의 선형 보간으로 만들어 질 수 있다. 도 11은 부드러운 영역의 해당 픽셀과 그 이웃 값을 나타낸다.Another texture-based feature to look at next is based on the similarity of neighbor pixel values in most images. Pixel values in the rest of the image, except the boundaries of the image, change smoothly. Subdividing soft areas into dark and light areas approximates the corresponding pixel values in each area.

Is given coefficient

And linear interpolation of surrounding pixel values. Fig. 11 shows corresponding pixels and their neighbor values in the soft area.

를 구하기 위해서 수학식 14와 같은 비선형 최적화 문제를 푼다.Satisfying the above conditions

To solve this problem, we solve a nonlinear optimization problem such as

를 구하면

을 구할 수 있다. 이 때,

의 평균과 표준 편차를 또 다른 특징점으로 둔다. RGB 각각에서 어두운 영역과 밝은 영역에 대하여 평균과 표준편차를 구할 수 있으므로 총 12개의 특징점을 구할 수 있다.
From here

If you find

Can be obtained. At this time,

Let the mean and standard deviation of be another feature point. Since the average and standard deviation can be obtained for the dark and bright areas of each RGB, a total of 12 feature points can be obtained.

Information theory based features

When outputting color images with halftone printing, each color channel contains inherent noise. Since the intensity of noise is different for each color channel, mutual information between the entropy of each color channel and the color channel can be regarded as an important characteristic point representing the characteristics of the object under test.

First, entropy may be expressed as in Equation 15.

Entropy is a measure of the degree of uncertainty in a given event. If the object under test is a counterfeit, the noise in the color channel will have an entropy value that is distinct from the original. Therefore, the entropy of each color channel can be seen as a characteristic point representing the inspection object.

Next, the mutual information is expressed as in Equation 16 below.

Mutual information is a measure of the dependency between event X and event Y. That is, if events X and Y are independent events, mutual information I (X; Y) has zero. However, if event Y always occurs when event X occurs, the mutual information value is equal to the entropy value of X.

If the test object is the original, the value of each color channel contains little noise, so the value of the color channel is dependent. As a result, the mutual information value becomes large. However, if the object to be inspected is a counterfeit, the inherent noise present in each color channel increases the independence between the channels and decreases the mutual information value. Therefore, mutual information can also be seen as a feature that represents an object to be inspected.

Summarizing the above, in this project, the values of entropy in each of R, G, and B channels and mutual information values between R, G, G, B, B, and R channels are used as information theory-based feature points.

Characteristic point based on network structure analysis

The color printer outputs the object by a dot printing method, unlike line drawing prints such as banknotes. Therefore, if a part of the object is scanned at a high resolution such that the halftone structure can be confirmed, a system capable of automatically detecting a forgery can be configured. Furthermore, the analysis of the scanned dot structure can identify the printer model that outputs the inspection object. A color printer typically uses four inks from CMYK to print the object. Each channel of CMYK is output at different angles to prevent moire effects, edge enhancement effects, and so on. FIG. 12 is a diagram of a gray image output by dot printing using a printer 3, one of commercially available printers, and separated into cyan, magenta, and yellow channels.

12, it can be seen that each color channel is output at different angles even though they have the same network structure. In addition, the network structure has various forms according to the manufacturer and printer model of the printer.

13 is a dot structure of a cyan region outputted to different printers.

In conclusion, we can infer that the angular information of the CMYK channel is a good feature that can be used to trace back the printer model. A Hough transform is applied to each channel to obtain angle information of each channel. The Hough transform may be expressed as in Equation 17.

The Hough transform can obtain most linear information in the image in the form of polar coordinates, that is, distance and angle from the origin. Since the halftone structure of the image scanned by the scanner is similar in most areas, the angle information of the extracted straight line also has a similar value. Therefore, if the histogram is constructed using the extracted angle information of the straight line, it is possible to construct a different type of histogram having a unique peak for each printer model.

14 shows the above-described halftone structure analysis process in detail.

14A illustrates an image output by a printer 5, which is one of commercially available printer models, and FIGS. 14B and 14C illustrate an image output by a printer 4. In each case, the figure on the left shows the halftone structure of the image and the straight lines representing it, and the middle figure shows the Hough transform process. Finally, the figure on the right is a histogram constructed from the angle information of the extracted straight lines. Comparing (a) and (b), it can be observed that the histogram constructed when different printers print similar images even if they are similar images. However, if you compare (c) and (d), you can see that even if the images are of different types, the histograms that are constructed have similar shapes when outputted to the same printer.

In contrast, linear prints, such as banknotes, do not have a regular printing pattern like a dot print, and thus do not generate histograms that form a specific distribution. Therefore, the histogram generated using Hough transform can be regarded as an important feature.

In the present invention, a histogram having 36 bins in three channels of Cyan, Magenta, and Yellow is constructed. Through the above, a total of 108 feature points are extracted.

Next, the techniques used in the machine learning unit 30 of the present invention will be described in more detail.

Feature Point Selection Technique

As described above, a number of feature points that can be extracted from the inspection object were described. However, using too many feature points can cause system overload. In fact, the performance of a system that identifies forgery of a security product is much lower than that of a typical PC.

Therefore, it is necessary to save resources such as computation and memory by selecting only the most appropriate feature points from a given set of candidate feature points. In order to find the most suitable feature point for counterfeit identification or model traceback among the given candidate feature points, an optimal feature point was selected using a sequential forward floating selection algorithm (SFFS). SFFS may be shown in the flowchart of FIG. 15 as a technique for constructing an optimal set of feature points by dynamically selecting or deleting feature points.

First, the terms used in the algorithm are defined as follows. m represents the total number of executions of the algorithm, and M represents the maximum number of executions. Z represents the set of selected feature points and U represents the set of all feature points. CR represents the classification performance calculated through 2-fold cross validation using the selected feature vector and SVM classifier. Finally, J is the classification result value when the classification performance is the best in step m. In steps 1 and 2, the SFFS algorithm selects one feature point that has the highest classification performance and adds it to Z. Next, we check if the total number of algorithms runs is greater than M. If m> M, the feature point selection history to date is examined and the set of feature points at the best classification result and the result of classification at that time are returned. In other cases, the process of dynamically removing Z from the selected feature points that is not necessary is performed in 4, 5, and 6 steps. Features that don't really help test performance continue to be removed from Z. If there are no more feature points to remove, go back to step 1 and add the required feature points.

16 shows the relationship between the number of feature points selected and the discrimination rate using SFFS. As can be seen from FIG. 16, when the number of selected feature points exceeds about 30, it may be confirmed that the classification performance is not significantly affected.

Through the process of SFFS, it is confirmed that the feature point based on discrete wavelet transform has a big influence on classification performance. In particular, it can be seen that after discrete wavelet transform, feature points between correlation coefficient-based feature points and subbands are mainly selected.

In the present invention, the feature point selector 31 selects some feature points from a plurality of feature points of the digital image by the SVM based method, and the learning model generator 32 selects the feature points. A learning model of the feature point vector set is generated from some feature points selected by (31).

The SVM technique used for learning in the feature point selector 31 and selecting a class in the SVM classifier 33 will be described in more detail below.

SVM Technique

는 주어진 특징점의 벡터이고,

는 그때 진위 여부를 나타낸다. After feature point selection and learning model generation, a technique is needed to identify forgery based on the feature points of a given test print. SVM is one of the most popular supervised learning methods for this purpose. To use an SVM classifier for forgery identification, you must first train it on a subset of the given input. Assuming that a given input is divided into two sets, it can be written as From here

Is the vector of the given feature points,

Indicates authenticity at that time.

로 나타내어진다. 각 집합에서

에 가장 가까운 점을 서포트벡터(support vector) 로 정의한다. SVM은 주어진 입력에서 서포트벡터와 두 집합을 나누는 다차원 평면 사이의 기하학적인 거리를 최대화 하는 것을 목표로 한다. 서포트벡터가 아래의 두 평면상에 존재 한다면 두 평면 사이의 거리는

이다. 또한, The multidimensional plane that divides these two sets

It is represented by In each assembly

Define the point closest to the support vector. SVM aims to maximize the geometric distance between the support vector and the multidimensional plane that divides the two sets at a given input. If the support vector is on two planes below, the distance between them

to be. Also,

17 is simply shown in FIG.

를 최대화 시키기 위해서는

를 최소화 시켜야 한다. 이상을 종합하면 수학식 19와 같이 최적화 문제를 유도할 수 있다.

To maximize

Should be minimized. In sum, the optimization problem can be derived as shown in Equation (19).

를

로 치환할 수 있고, 이는

와

를 찾기 위한 수학식 20과 같은 전형적인 이차 비선형 최적화 문제가 된다.In the present invention without losing generality

To

Which can be substituted by

Wow

It is a typical quadratic nonlinear optimization problem, such as (20).

와

를 구할 수 있다. 이상의 과정에서 우리는 주어진 입력의 진위 여부를 판별할 수 있는 SVM분류기를 학습시켰다. 이후 주어진 입력

에 대하여

이면

는 첫 번째 클래스로 분류되며,

이면 두 번째 클래스로 분류된다. Solving a given problem in the dual realm

Wow

Can be obtained. In the above process, we have learned an SVM classifier that can determine the authenticity of a given input. Since given input

about

If

Is classified as the first class,

Is classified as the second class.

에 대하여 가장 높은 결과 값을 보이는 SVM분류기의 집합으로 분류한다. If we need to classify a given problem as a class with more than the number of image channels, K> 2, we will learn K binary SVM classification problems that divide the set of one set with all other sets. Since given input

Classify the set of SVM classifiers with the highest result.

In the above, we looked at the SVM classifier when linear separation is possible. However, input data is not always linearly separable. Even when the linear separation is impossible, the SVM classifier is composed by introducing the concept of Slack variable and Kernel function to minimize the error rate of the classification function. Slack variables are used to minimize errors in the linear classifier by adding the appropriate error values to the multidimensional plane in advance, even though the two classes are not completely linearly divided. The SVM classifier to which the slack variable is applied is shown in Equation 21.

However, slack variables alone cannot completely separate input data that cannot be linearly separated. Therefore, we introduce the concept of a kernel function that maps input data of nonlinear separation into a form capable of linear separation. The concept of a kernel function is shown in FIG.

Through the kernel technique, it is possible to confirm that two-dimensional data that cannot be linearly separated into three-dimensional map and change into a form capable of linear separation accordingly. In the present invention, nonlinear input data is linearly mapped using a radial basis function (RWB) kernel as shown in Equation (22).

로 나타나는 2개의 변수가 추가된다. 이는 교차검증(cross-validation)을 통해 최적의 분류 성능을 나타내는 값으로 결정된다.
The introduction of Slack variables and kernel functions

Two variables are indicated by. This is determined by a value representing the optimal classification performance through cross-validation.

Finally, the counterfeit identification unit 40 can determine whether the forgery of the test print successfully by the class selected by the SVM classifier 33 as described above.

100: security print counterfeit identification system
110: print learning device 120: print identification device
10: image conversion unit 20: feature point extraction unit
30: machine learning part 40: counterfeit identification part
50: first image converter 60: first feature point extractor
70: second image conversion unit 80: second feature point extraction unit
31: feature point selector 32: learning model generator
33: SVM Classifier

Claims (19)

In security printed counterfeit identification system using digital forensic technology,
An image converting unit converting the printed matter to generate a digital image;
A feature point extraction unit for extracting feature points of the digital image;
A machine learning unit for performing machine learning using the feature points; And
Security forgery identification system using a digital forensic technology, including; forgery identification unit for identifying whether the forgery of the print from the machine learning results.
In the secure forgery identification system using digital forensic technology, including a print learning device and a print identification device,
The printed matter learning device,
A first image conversion unit converting the original prints and the fake prints into digital images to generate original digital images and fake digital images;
A first feature point extracting unit for extracting feature points of the original digital image and the counterfeit original digital image;
A feature point selector for selecting some feature points by a SVM (Support Vector Machine) method using a plurality of feature points of the original digital image and the counterfeit digital image; And
And a learning model generator for generating a learning model of the feature point vector set based on some feature points selected by the feature point selector.
The printed matter identification device,
A second image converting unit converting the test print into a digital image to generate a test print digital image;
A second feature point extracting unit extracting feature points of the test print digital image;
An SVM classifier for selecting a class to be included in the test print by using feature points of the test print digital image based on the generated learning model; And
Security forgery forgery identification system using a digital forensic (Forensic) technology comprising a; forgery identification unit for identifying whether the forgery of the test print from the class of the SVM classifier.
The method of claim 1,
The image conversion unit,
A first image conversion unit converting the original print and the counterfeit print out of the prints into a digital image to generate an original digital image and a counterfeit print digital image; And
And a second image converting unit converting the test print out of the prints into a digital image to generate a test print digital image.
The feature point extraction unit,
A first feature point extracting unit for extracting feature points of the original digital image and the counterfeit original digital image; And
And a second feature point extracting unit extracting a feature point of the test print digital image. The printed matter forgery identification system of claim 1, further comprising a digital forensic technology.
The method of claim 3, wherein the machine learning unit,
A feature point selector for selecting some feature points by an SVM (Support Vector Machine) method using a plurality of feature points from the original digital image and the fake digital image;
A learning model generator for generating a learning model of a feature point vector set based on some feature points selected by the feature point selector; And
Security printed material forgery identification system using a digital forensic technology, characterized in that it includes; SVM classifier for selecting a class to be included in the test prints using the feature points of the test printout digital image based on the generated learning model .
The method of claim 4, wherein
The feature point selector is a security forgery identification system using a digital forensic technology, characterized in that for selecting some feature points using a sequential forward flow selection algorithm.
The method of claim 3,
Wherein the first feature point extractor or the second feature point extractor constitutes a feature point vector from the extracted feature points,
The feature point,
Feature of discrete wavelet transform;
Feature point based on image quality measurement tool;
Texture-based feature points;
Information theory-based features; And
Security point forgery identification system using digital forensic technology, including;
The method of claim 6, wherein the discrete wavelet transform-based feature point,
Feature points due to statistical features in high frequency subbands;
Feature points due to statistical features between high frequency subbands; And
Characterized by the correlation coefficient between the color channels of the high frequency subband; including,
The high frequency subband has a high frequency characteristic in any one or more of the horizontal direction and the vertical direction among four subbands determined by performing a low pass filter, a high pass filter, and a downsampling process in the horizontal and vertical directions of the image. Security printed counterfeit identification system using digital forensic technology, characterized in that it comprises three sub-bands.
The method of claim 7, wherein the characteristic point due to statistical features in the high frequency subband,
Security printed material forgery identification system using digital forensic technology, characterized in that it comprises a standard deviation, distortion and kurtosis of each high frequency subband after the discrete wavelet transform.
The method of claim 7, wherein the feature point by the statistical feature between the high frequency subband,
Security forgery identification system using digital forensic technology, characterized in that it includes the average, variance, kurtosis, and skewness of the error of the approximate value of each subband and the subband value after the discrete wavelet transform.
The method of claim 7, wherein the feature point by the correlation coefficient between the color channels,
Between R and G, between R and B, between G and B, between C and M, between C and Y, between C and K, between M and Y, and between M and Y determined by performing discrete wavelet transform for each color channel. A security printed material forgery identification system using digital forensic technology, characterized in that it is a correlation coefficient for high frequency subbands between K and Y and K.
The method of claim 6, wherein the feature-based feature point,
In the spatial domain between any input image and the image from which noise is removed from the image input by the low frequency filter, the difference value of the pixel, the difference value of the coefficient in the frequency domain, the correlation coefficient of the two inputs, and the cognitive visual model are applied. Security printed material forgery identification system using digital forensic technology, characterized in that the extracted feature point based on any one or more of the difference value of the pixel.
The method of claim 11, wherein the low frequency filter,
Security forgery identification system using digital forensic technology, characterized in that it comprises an averaging filter, Gaussian filter, median filter, and Wiener filter.
The method of claim 6, wherein the texture-based feature point,
Security imprint forgery identification system using digital forensic technology, characterized by including three statistical characteristics of energy, contrast and similar obtained using Gray Level Co-occurrence Matrices (GLCM).
The method of claim 13, wherein the texture-based feature point,
Security forgery identification system using digital forensic technology, characterized in that it further comprises the average and standard deviation of the difference between the arbitrary pixel value and the pixel approximation value.
According to claim 6, The information theory-based feature point,
Security imprint forgery identification system using digital forensic technology, characterized in that it includes entropy values in each of R, G, and B channels, and mutual information values between R, G, G, B, B, and R channels. .
The method of claim 6, wherein the feature point based on the network structure analysis,
A digital forensic, characterized by calculating the straight line information present in the image through the Hough transform in the form of distance and angle from the origin and using the calculated angle information of the straight line. Security imitation identification system using technology.
The method of claim 2, wherein the SVM classifier,
Security forgery identification system using digital forensic technology characterized by applying Slack variable and Kernel function to minimize error rate of classification function.
The method of claim 17, wherein the kernel function (Kernel),
Security printed material forgery identification system using digital forensic technology, characterized in that the RBF (Radial Basis Function) Kernel.
In the security print forgery identification method by the security forgery identification system using digital forensic technology including an image conversion unit, feature point extraction unit, machine learning unit and forgery identification unit,
An original digitization step of generating an original digital image by converting the original printout into a digital image;
A counterfeit digitizing step of generating a counterfeit digital image by converting the counterfeit printed matter into a digital image by the image converting unit;
An original feature point extraction step of extracting a feature point of the original digital image by the feature point extraction unit;
A counterfeit feature feature extraction step of extracting a feature point of the counterfeit bone digital image by the feature point extractor;
A feature point selecting step of the machine learning unit selecting some feature points from the feature points of the original digital image and the feature points of the fake digital image by a SVM (Support Vector Machine) method;
A learning model generation step of generating, by the machine learning unit, a learning model of a feature point vector set based on some feature points selected in the feature point selection step;
A test print digitizing step of generating a test print digital image by converting the test print into a digital image by the image conversion unit;
A test print feature point extracting step of extracting a feature point from the test point digital image by the feature point extracting unit;
An SVM classification step of selecting, by the machine learning unit, a class to include the test print using the feature points extracted in the test print feature point extraction step based on the learning model generated in the learning model generation step; And
Security forgery identification method using a digital forensic technology comprising a; forgery identification step of identifying whether the forgery of the test print by the class selected in the SVM classification step in the forgery identification unit.

Images