KR101124488B1 - Color laser printer identification method through noise characteristics analysis - Google Patents

Abstract

A method for discriminating a color laser printer through a Wiener filter and a contrast matrix is provided. After scanning a digital image printed by a color laser printer, the RGB domain area is converted into a CMY area. An image, from which noise is removed, is generated through a Wiener filter on the image represented by the CMY region. A noise component is extracted based on the difference between the CMY image and the image from which the noise is removed. A matrix value is calculated by performing contrast-contrast matrix transformation on the extracted noise component. A plurality of feature points with similarity, contrast, sum, correlation coefficient, and covariance values are calculated for the matrix values. The plurality of feature points are combined into one feature vector and input as features of a support vector machine classifier to determine the color laser printer.

Description

Color laser printer identification method through noise characteristics analysis

The present invention relates to a color laser printer, and more particularly, to a method for discriminating a color laser printer corresponding to an image acquisition device among digital image forensic technology.

Due to the popularization of computers, our society can purchase high-performance computers and peripherals at low prices. Also, due to the development of various softwares, even beginners can use the computer to easily perform complicated tasks. This has some positive aspects but also some negative ones. A typical example is forgery and forgery. 1 is a view showing the status and detection rate of forgery and forgery crime by year. As shown in FIG. 1, it can be seen that as the years go by, various kinds of forgery and alteration crimes increase. This is because forgery and forgery are easy due to the development of image editing programs such as Photoshop and Illustrator. In addition, as forgery and alteration techniques become more sophisticated, the probability of discriminating forgery and alteration by the general public is very low (1%).

Due to the increase in image forgery and modulation, image reliability is becoming more important. Digital image forensic technology is being actively researched to enhance the reliability of such images. Digital image forensic technology is a technology to find out the inherent characteristics hidden in the image to identify the device that produced the image or to determine whether the image is forged. The research is divided into the image acquisition equipment identification and the image forgery detection technology.

Image acquisition equipment identification technology is a technique for determining what equipment used to acquire an image from a given image. It aims to determine the exact model name level as well as the manufacturer of the equipment used to acquire the image using only the given image.

Image forgery detection technology is a technique for determining whether a given image has undergone a forgery / modulation step. Forgery and alteration of digital images leave no visible clues, but change the statistical properties of the image. Therefore, it is possible to find out whether the forgery or alteration is found by finding the changed statistical characteristics.

The printer discrimination method technology studied so far is mainly a technology for discriminating monochrome laser printers. However, as the penetration rate of color laser printers increases, the necessity of discriminating technology suitable for color laser printers is emerging.

There are two methods of identifying a printer using an output document. There are two methods, a passive method and an active method. The passive method identifies printers using intrinsic signatures by looking for invisible features on the output that are inevitably generated due to physical properties within the printer and finding the printer that printed them. Active methods identify printers by using extrinsic signatures by directly modifying parameters during the printer's printing process and inserting certain invisible information directly into the output. The research on printer discrimination technology is being conducted according to the passive method rather than the active method because it can cause consumers to insert specific information into the document using the active method.

Printer steganography is a kind of active printer discrimination method, and is a technique of outputting yellow dots on printed paper to prevent information from being illegally used or tampered with by a third user. These yellow spots are hard to see by sight, and can be decrypted using high-resolution scanning equipment or special rays of light. This yellow dot contains various information such as the printer's unique serial information and the date printed.

2 and 3 are diagrams showing examples of printer steganography. In the image of FIG. 2, yellow dots are displayed to indicate the output time, date, equipment number, and the like. The image of the printed document using printer steganography is shown in the right image of FIG. 3. can see. This technology has been repulsed by users because it is a serious privacy breach that is embedded in a yellow printer by encrypting private information in a document printed on a laser printer. In addition, even if it is inconspicuous, the printed image may be printed once, so that the image quality and the quality of the printed product may be degraded. Therefore, recently, a method of researching digital forensic technology that is not machine dependent has been studied.

During the printing process of laser printers, light and dark streaks occur periodically along the printing direction due to the quasi-cyclic fluctuations caused by the angular velocity becoming inconsistent due to the reverse rotation of the gears rotating the optical photoconductor drum. The stripes on the document reflect the mechanical characteristics of the printer. Delp et al. Proposed an algorithm for discriminating black and white laser printers using these periodic characteristics. 4 is a view showing the fringes of the output of the output by the monochrome laser printer.

In general, a method of obtaining a stripe period using a halftone image output may be performed by projecting in a horizontal direction and then using a binary Fourier transform, as in the method of obtaining a stripe period of a printer. The printer discrimination method finds a printer having a stripe period that matches the stripe period after obtaining the stripe period of the halftone image output which is not printed with which printer by the above method. This stripe period is easy to apply to halftone image output and results are good, but it is difficult to apply to plain text output that is not halftone image output.

Although the documents printed with a laser printer may look the same at a distance, the magnified view shows features distinguished by the quality of the printer. 5 shows an example of the distinguishing features depending on the printer quality. In the printer discrimination method using distance conversion, a distance conversion algorithm is used to measure the distance between text image matching and text existing in a printed document. As shown in FIG. 5, the characteristics of the output result that are slightly different for each printer are analyzed and used for printer discrimination.

Character image matching is generally based on distance transformation using binary images. The binary image I (x, y) has only two values, 0 and 1, where 0 is the letter pixel and 1 is the background pixel. I _d (x, y) represents the distance transformation of the image, which is the distance between the pixel (x, y) of the text and the background pixel closest to this pixel (x, y). The calculation of the distance transform is performed using a two-pass algorithm. The distance difference between the two binary images after the distance conversion may be calculated by Equation 1 below.

Where H and W represent the height and width of the image, respectively, and N _f is the number of black pixels in the image (x, y). g _d (x, y) is the distance transform of the g (x, y) image. Since different matching distances can be obtained according to the matching directions, two matching directions D _fg and D _gf are calculated and used in combination.

Characters printed by previously known printers are recognized using the OCR function, and the recognized characters are registered in the database. After that, when an image comes in for printer identification, a character is recognized and the distance difference value is calculated through a distance conversion method by comparing with a database. With respect to the distance difference value thus calculated, a printer including characters in the case of having the minimum distance difference is determined as the printer used to print the document. The structure of such a printer identification system is shown in FIG. 6 is a view for explaining a printer identification method using a distance conversion.

The printer discrimination technique using discrete wavelet transform has been suggested by Choi as a printer discrimination method, focusing on the fact that the color reconstruction method differs slightly when a digital image is output as an actual document for each color laser printer. Unlike conventional black and white laser printer discrimination technology, this is a technology for discriminating a color laser printer using color information of a color document.

A printer discrimination process using discrete wavelet transform is illustrated in FIG. 7. Outputs printed by several color printers are scanned by one scanner, and then discrete wavelet transform is performed for each of R, G, and B areas in the RGB area, which is the basic color area.

FIG. 8 is a diagram illustrating an HH region formed after a discrete wavelet transform. Of the four regions created after the discrete wavelet transform, HH in FIG. The noise is statistically analyzed to extract the feature points. Find the standard deviation, skewness, kurtosis, R and G, G and B, and R and B band covariances and correlation coefficients in the HH region with noise components among the discrete wavelet transforms. Set them as one feature point vector. A total of 15 feature points such as standard deviation, skewness, kurtosis, covariance, and correlation coefficient can be extracted from the RGB domain. The equations for each feature point are summarized in Table 1. That is, Table 1 shows the feature points of the color laser printer discrimination method using discrete wavelet transform. This process is the same for the YCbCr domain. Thus, a total of 30 feature points extracted from the two regions are combined into one vector. One feature point vector is generated for each image, and then half of the feature point vectors generated using the support vector machine are used for learning, and the other half is used to determine the color laser printer outputting the document.

Feature Point Equation Standard Deviation

Kurtosis

Dwarf

Covariance

Correlation coefficient

SUMMARY OF THE INVENTION The present invention has been made to solve a conventional problem, and an object thereof is to provide a method for determining a color laser printer through a Wiener filter and a contrast matrix.

In order to achieve the above object, the color laser printer discrimination method according to the present invention comprises the steps of (i) scanning the digital image printed by the color laser printer and converting the RGB domain region to the CMY region; (ii) generating an image from which noise is removed through a Wiener filter on the image represented by the CMY region; (iii) extracting a noise component based on a difference between the CMY image and the image from which the noise is removed; (iv) calculating a matrix value by performing contrast-contrast matrix transformation on the extracted noise component; (v) calculating a plurality of feature points with similarity, contrast, sum, correlation coefficient, and covariance values for the matrix values; And (vi) combining the plurality of feature points into one feature vector and inputting the feature as a feature of the support vector machine classifier to determine the color laser printer.

The present invention proposes a new discrimination method for a color laser printer currently being developed, and achieves an excellent discrimination rate compared to the existing studies. The method of the present invention showed an accuracy of 97.6% in determining a manufacturer of a color digital printing machine, and an 84.5% accuracy in determining a model of the same manufacturer. According to the present invention, a color laser printer discrimination technology, which is a field of digital forensic technology, can know which printer outputs the digital image output from the printer. Counterfeit documents may be easily identified.

1 is a view showing the status and detection rate of forgery / forgery crime by year.
2 and 3 are diagrams showing examples of printer steganography.
4 is a view showing the fringes of the output of the output by the monochrome laser printer.
5 is a diagram showing an example of distinguished features according to printer quality.
6 is a view for explaining a printer identification method using a distance conversion.
7 is a diagram illustrating a printer discrimination process using discrete wavelet transform.
FIG. 8 is a diagram illustrating an HH region formed after a discrete wavelet transform.
9 is a view for explaining a color laser printer discrimination method according to an embodiment of the present invention.
10 to 13 are diagrams showing examples represented by the CMY region.
14 illustrates a CMY image.
15 is a view illustrating an image from which noise passed through a Wiener filter is removed.
16 to 19 illustrate noise of each manufacturer using the Wiener filter.
20 is a diagram illustrating an example of contrast degree simultaneous matrix directionality according to an embodiment of the present invention.
21 is a diagram illustrating an example of binary classification and mapping using SVM.
22 is a diagram illustrating an example of an image used in an experiment according to an embodiment of the present invention.
Fig. 23 is a diagram showing the result of the accuracy analysis of the color laser printer manufacturer determination.
24 is a diagram illustrating a result of analyzing accuracy of discrimination by model of the same color laser printer of the same manufacturer.

Hereinafter, a color laser printer discrimination method according to an embodiment of the present invention will be described with reference to the accompanying drawings.

9 is a view for explaining a color laser printer discrimination method according to an embodiment of the present invention. The color laser printer discrimination method according to the present invention is largely divided into a training process and a discrimination process as shown in FIG. 9. In the training process, the characteristics of each printer are extracted by using the images printed by the printer to be discriminated, and the training is performed by inputting the parameters of the learning-based data classifier. In the discriminant process, when arbitrary data enters, the feature is extracted, inputted into the learning-based data classifier, and the printer determines which printer is output.

Since the performance of the color laser printer discrimination algorithm is greatly influenced by feature extraction to uniquely identify the printer, the present invention has been conducted on various feature points in order to extract unique features present in the printer image. The method of representing an optimal result and accurately determining the unique features of each printer is the same as the method of extracting feature values from the feature point extractor of FIG. 9.

Hereinafter, a method of extracting a noise using a Wiener filter according to the present invention, a feature extraction method using a contrast-concurrency matrix, and a learning-based data classifier will be described.

To extract the unique features of color laser printers, first scan a digital image printed on a color laser printer with a scanner (not shown), convert it from an RGB image to a CMY image, and then use a Wiener filter (not shown). Create a noise-free image. Only noise was extracted from the difference between the original image and the image from which the noise was removed. To improve the classification accuracy, we compute the contrast coherence matrix in four directions of 0 °, 45 °, 90 °, and 135 ° for the extracted noise, and then compare the contrast, correlation, sum, similarity, Covariance was calculated and used as the feature point.

The CMY region was used for the domain region to be used. When a printer outputs a document, it converts the RGB area used for representing the document on the computer to the CMY area used for the printing process and then outputs the document. Therefore, when the scan is converted again from the RGB area to the CMY area, the noise generated when output to the printer is less damaged, and thus the printer-specific difference is more visible than other domains. CMY has three colors: Cyan, Magenta, and Yellow. An example represented by the CMY region is shown in FIGS. 10 to 13. 10 shows a CMY image, a C image in FIG. 11, an M image in FIG. 12, and a Y image in FIG. 13.

The Wiener filter is useful for removing abnormal noise. 14 and 15 are diagrams illustrating noises removed through a CMY image and a Wiener filter, that is, an image transformed into a CMY region, respectively, through the Wiener filter. FIG. 15 shows that the image is softened by removing noise compared to FIG. 14. The image transformed into the CMY area is removed by using the Wiener filter to remove unintended or popped values to create a new image. By calculating the difference between the original image and the noise-removed image, we can find the noises that were inadvertently introduced into the original image. Since these noises are different for each manufacturer as shown in FIGS. 16 to 19, the noises may be a standard for identifying a printer. 16 to 19 are noises of each manufacturer using the Wiener filter, FIG. 16 shows noise of Xerox printer, FIG. 17 shows noise of HP printer, FIG. 18 shows noise of Canon printer, and FIG. 19 shows noise of Konica printer. Represent each.

The basic concept for pixel-based statistical texture image generation is organized by the Gray Level Co-occurrence Matrix proposed by Haralick et al. Application studies of texture images have shown in many case studies that texture images are useful for qualitative classification and improvement of quantitative classification accuracy. However, although texture images are useful, little research has been done on the selection criteria for texture images.

20 is an example of contrast and concurrency matrix directionality according to an embodiment of the present invention. Contrast matrix is a method of expressing the relationship between brightness values between pixels in an image. The coordinate value is obtained by using values of two adjacent pixels as a coordinate value as a distance between a direction and two points as shown in FIG. 20. It is a matrix of frequent occurrences.

In the present invention, in order to take advantage of the contrast co-occurrence matrix which improves the classification accuracy, the noise co-occurrence matrix was calculated for four directions of 0 °, 45 °, 90 °, and 135 ° using the noise obtained by the Wiener filter. . As shown in Table 2, each feature point was extracted using statistical characteristics of similarity, contrast, sum, correlation coefficient, and covariance, as shown in Table 2. Table 2 describes the feature points for the intensity co-occurrence matrix.

Feature Point Equation Explanation Similarity

Express a uniform degree between each pixel in the matrix contrast

Expresses the difference in contrast between pixels Sum

Express the sum of the pixels Correlation coefficient

Express correlation between pixels Covariance

Expresses the degree to which two pixels are related to the entire image

Statistical characteristics include the similarity, contrast, sum, and correlation variance for each of the C, M, and Y regions obtained by performing concurrency matrix transformation for the four directions, and between C and M, M, Y, Y, and C regions. There is covariance. The 60 feature points extracted in this way are combined into one vector and used as training and discrimination process data for identifying a color laser printer in a learning-based data classifier.

As a learning-based data classifier, Support Vector Machine (SVM) is a learning algorithm developed and proposed by statistician Vladimir Vapnik in 1995 and developed for binary classification as shown in FIG. 21A. Binary classification is the use of collected training data to find the optimal linear decision surface based on the concept of structural risk minimization. SVMs currently use bioinformatics, character recognition, handwriting recognition, It has been successfully applied in various fields such as face and object recognition.

21 shows an example of binary classification (left) and mapping (right) using SVM. As shown in the left figure of FIG. 21, the SVM maps an input vector into a high-dimensional feature space to maximize a margin among possible hyperplanes that can be classified into different classes. hyperplane). The training data closest to the maximum margin hyperplane is called the support vector, and the larger the margin value between the hyperplanes containing it, the better the classification performance.

SVM has the advantage of easy mathematical analysis because it converts a problem with a very complex separation boundary into a simple problem that can be used with a linear discriminant function. Also, the classification is good even with a small amount of learning materials. SVM has superior generalization performance based on the concept of structural risk minimization, which subdivides the whole group into subgroups and then minimizes the empirical risk for this group, compared to the existing principle based on empirical risk minimization that minimizes learning errors. It also provides various kernels to solve complex classification problems such as nonlinear separation problems.

Various software such as SVM light, LIBSVM and SVM Torch are publicly available and relatively easier to use than other machine learning methods. When using SVM, all you have to do is select the kernel and the corresponding kernel variables. In addition, new patterns can be used to dynamically update the model during training.

In the present invention, LIBSVM is used to enable multi-class classification among support vector machines. LIBSVM adds a penalty for unbalanced classes, which minimizes sorting errors in the user's hyperplanarization to ensure that multiple groups are clearly classified.

Kernel functions are theoretically possible, but instead of mapping functions that are difficult to solve, they move the raw data into higher-dimensional spaces to create a linearly separable set of input data within a specific space. The kernel function is defined as in Equation 2 below.

In general, there are four types of kernel functions: linear, polynomial, radial basis function, and sigmoid, as shown in Table 3. Where γ, r, and d refer to parameters in each kernel. Table 3 shows representative kernel functions of the support vector machine.

Kernel functions Equation Linear

Polynomial

RBF

Sigmoid

In the present invention, a Radial Basis Function (RBF) kernel is used. The RBF kernel is very useful for solving the problem that linear separation is impossible because it moves nonlinearly the input vector into high-dimensional feature space. The RBF kernel is affected by the predictive performance of the SVM based on C, which is the upper tolerance of Lagrange multipliers, and γ, the kernel parameter. Therefore, overfitting or underfitting can occur if two parameters are not properly selected.

Experiment result

In this experiment, a total of 2,597 images were printed out, using 371 images from 7 printers (HP, Canon, Xerox DCC 400, Xerox DCC 450, Xerox DCC 4550, Xerox DCC 6540, Konica). In order to acquire the image data, 371 images were printed using seven color laser printers, scanned at 300 dpi with an EPSON Perfection 2400 scanner, and divided into 256 × 256 size for the experiment. 22 is a part of an image set used in an experiment according to an embodiment of the present invention.

In order to analyze the performance of the algorithm according to the present invention, two experiments were performed for the printer manufacturer and the model for the same manufacturer. In order to discriminate by color laser printer manufacturer, experiments were conducted using a total of 1,484 images of 371 prints printed by four printers of HP, Canon, Xerox DCC 450 and Konica, and Xerox DCC 400 and Xerox DCC Experiments were performed using 1,484 images of 371 printers, each of which was printed on four printers of 450, Xerox DCC 5540, and Xerox DCC 6540 models.

The algorithm according to the present invention and the Choi algorithm using Discrete Wavelet Transform (DWT) were tested using a support vector machine classifier, respectively. For the objectivity and accuracy of performance discrimination, both experiments used 371 prints of each manufacturer and randomly selected 1,484 printed images. 742 were used as training data and the remaining 742 were used as test data. In addition, the cross-validation technique was used to minimize the error of the experiment. Two experiments were conducted using the test data as the training data and the test data as the test data by changing the randomly selected training data and the test data. The final results were calculated by averaging the two results.

Performance analysis by manufacturer

Discriminant performance of each manufacturer was analyzed using the proposed algorithm and Choi's algorithm using DWT under the same experimental environment. The manufacturers of color laser printers used in the experiments are summarized in Table 4. That is, Table 4 describes the color laser printers used in the printer discrimination experiment by manufacturer.

label Printer model Xerox Xerox DCC 400 HP HP 4650 cannon Canon iR C2620 Konica Konica Minolta bizhub C250

Table 5 shows the results of discriminating color laser printer manufacturers using the algorithm using DWT and the proposed algorithm.

Xerox HP cannon Konica Invention How to use DWT Invention How to use DWT Invention How to use DWT Invention How to use DWT Xerox 98.9% 98.3% 0.3% 0.7% 0.3% 0.5% 0.5% 0.6% HP 0.3% 7.0% 98.9% 91.4% 0.3% 1.6% 0.5% 0 % cannon 0 % 15.6% 0.8% 1.1% 98.4% 76.3% 0.8% 7.0% Konica 1.6% 2.2% 1.3% 0.2% 0.5% 0.2% 96.5% 97.3%

23 shows the result of the accuracy analysis of the color laser printer manufacturer determination. A comparison of the accuracy of the discrimination rate for the manufacturer is shown in FIG. 23. As a result of the experiment, the classification result between printers of the proposed method is 98.2%, which is 7% higher than the classification result of the algorithm using DWT. In addition, Canon printer showed higher discrimination rate than the algorithm using DWT.

Performance analysis by model of same manufacturer

The discriminant performance of each model was analyzed using the proposed algorithm and Choi's algorithm using DWT under the same experimental environment. The model names of the color laser printers used in the experiment are summarized in Table 6. That is, Table 6 describes the color laser printers used in the printer discrimination experiment for each model.

label Printer model DCC400 Xerox DCC 400 DCC450 Xerox DCC 450 DCC5560 Xerox DCC 5560 DCC6540 Xerox DCC 6540

Table 7 shows the result of determining the model name of the same color laser printer model, that is, the model name of the same color laser manufacturer using the algorithm using the DWT and the proposed algorithm.

label DCC400 DCC450 DCC5560 DCC6540 Invention How to use DWT Invention How to use DWT Invention How to use DWT Invention How to use DWT DCC400 96.5% 92.2% 1.9% One % 0.5% 3.2% One % 3.5% DCC450 0.8% 0.8% 94.6% 72% 1.3% 14.6% 3.2% 12.7% DCC5560 0.5% 6% 0.5% 4.6% 76.8% 61.5% 22.1% 28% DCC6540 0.3% 7% 0.5% 11.3% 29.1% 35.6% 70% 46%

A comparison of the accuracy of the discrimination rate for a manufacturer is shown in FIG. 24. That is, FIG. 24 illustrates a result of analyzing accuracy of discrimination by model of the same color laser printer. As a result of the experiment, the classification result between printers of the method of the present invention was 84.5%, which was about 16% higher than 67.93% of the classification result of the algorithm using DWT. In addition, the DCC450, DCC5560, and DCC6540 models show higher discrimination rate than the algorithm using DWT.

Claims (4)

(i) converting an RGB domain region into a CMY region after scanning a digital image printed by a color laser printer;
(ii) generating an image from which noise is removed through a Wiener filter on the image represented by the CMY region;
(iii) extracting a noise component based on a difference between the CMY image and the image from which the noise is removed;
(iv) calculating a matrix value by performing contrast-contrast matrix transformation on the extracted noise component;
(v) calculating a plurality of feature points with similarity, contrast, sum, correlation coefficient, and covariance values for the matrix values; And
(vi) combining the plurality of feature points into a feature vector and inputting the feature as a feature of a support vector machine classifier to determine the color laser printer;
In the step (v), the plurality of feature points include C (cyan) obtained by performing a contrast-contrast matrix transformation on the extracted noise components in four directions of 0 °, 45 °, 90 °, and 135 °, 60 kinds of similarity, contrast, sum, correlation variance for each region of M (magenta) and Y (yellow) and covariance between the C region and the M region, the M region and the Y region, and the Y and C region Color laser printer discrimination method that is a feature point. delete The method of claim 1, wherein K (χ _i , χ _j ) = φ (χ _i ) ^T φ (χ _j ), which is theoretically possible but replaces the mapping function that is difficult to solve in practice. A method of determining a color laser printer applying a kernel function to a support vector machine classifier of step (vi), which moves to create a linearly separable set of input data within a specific space. The method of claim 3, wherein the kernel function is

And a circular reference function kernel that solves the problem that linear separation is impossible by non-linearly moving the input vector into a high-dimensional feature space.

Images