Method for creating learning model for authenticity discrimination, authenticity discrimination method, and software

Abstract

To provide a method for creating a learning model for authenticity discrimination, an authenticity discrimination method, and software that are unaffected by noise from non-dot-point regions and that impose no limitations on a reading device.SOLUTION: A method for creating a machine learning model comprises: a first image data acquisition step of acquiring respective image data of images of a plurality of genuine articles and images of a plurality of non-genuine articles; a first frequency image conversion step of performing a frequency conversion on the image data of the plurality of genuine articles and the image data of the plurality of non-genuine articles to create frequency-converted image data of genuine articles and frequency-converted image data of non-genuine articles; and a learning model generation step of executing a machine learning algorithm on the frequency-converted image data of genuine articles and the frequency-converted image data of non-genuine articles to generate a learning model by deriving, based on the machine learning algorithm, the similarity for authenticity for the frequency-converted image data of genuine articles and the frequency-converted image data of non-genuine articles.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a method for deriving the similarity of authenticity of security printed materials such as banknotes, passports, securities, identification cards, cards, and travel tickets that require anti-counterfeiting effects, and to creating a learning model for authenticity determination, a method for authenticity determination, and software. [Background technology]

[0002] Security printed materials such as banknotes, passports, securities, and identification cards require anti-counterfeiting technology to prevent duplication and counterfeiting. Among the various anti-counterfeiting technologies, there is a particular need for anti-counterfeiting technology, such as watermarks and holograms, which do not require tools and can be used by anyone who holds a printed material to determine its authenticity.

[0003] As one example, a technology has been disclosed in which a surface image of a printing medium made up of halftone dots is optically acquired, the image information is analyzed using a frequency analysis method such as a Fourier transform, and position information of the frequency components calculated is generated that corresponds to the spacing L of the halftone dots forming the image and the inclination angle θ of the straight line formed by connecting the halftone dots, and this information is compared with position information of a genuine medium that has been stored in advance in a storage means to determine authenticity (see, for example, Patent Document 1). [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Laid-Open No. 2005-31802 Summary of the Invention [Problem to be solved by the invention]

[0005] However, the technology described in Patent Document 1 involves optically acquiring a surface image of a printing medium consisting only of halftone dot areas and analyzing the image information using a frequency analysis technique such as a Fourier transform. In printed matter that has not only halftone dot areas but also non-halftone dot areas formed by lines, the influence of noise from the non-halftone dot areas is significant, making it difficult to extract positional information of frequency components relative to the printing medium.

[0006] Furthermore, when determining authenticity, it is necessary to use pre-stored position information for genuine media, but the magnification and reading angle (distance between the camera and the subject) of the reading device that captures the image of the surface of the printed media must be determined in advance, which poses the problem of limited reading devices.

[0007] The present invention aims to solve the above problems and provides a method for determining authenticity that is free from the influence of noise in non-dot areas and is not limited by the type of reading device. [Means for solving the problem]

[0008] The present invention is a method for creating a learning model for authenticity discrimination, characterized by comprising: a first image data acquisition step of acquiring image data for each of a plurality of images of genuine products and a plurality of images of non-genuine products; a first frequency image conversion step of performing frequency conversion on the image data of the plurality of genuine products and the image data of the plurality of non-genuine products to create frequency-converted image data of genuine products and frequency-converted image data of non-genuine products; and a learning model generation step of executing a machine learning algorithm on the frequency-converted image data of genuine products and the frequency-converted image data of non-genuine products, and deriving the similarity of authenticity between the frequency-converted image data of genuine products and the frequency-converted image data of non-genuine products based on the machine learning algorithm to generate a learning model.

[0009] The present invention is a method for determining whether an item is genuine or not, using a learning model created by a learning model creation method, and is characterized by comprising: a second image data acquisition step for acquiring image data of at least a portion of the item to be determined; a second frequency image conversion step for frequency-converting the image data of the item to be determined to create frequency-converted image data of the item to be determined; a similarity derivation step for inputting the frequency-converted image data of the item to be determined into the learning model and deriving the similarity of the frequency-converted image data; and a discrimination step for determining whether the item to be determined is genuine or not based on the results of the similarity derivation step.

[0010] The present invention is a method for creating a learning model for authenticity discrimination, characterized in that the authentic product is a printed matter or paper on which elements having a predetermined regularity are applied.

[0011] The present invention is software that causes a computer to execute a learning model creation method.

[0012] The present invention is software that causes a computer to execute a method for determining whether a document is genuine or not. [Effects of the Invention]

[0013] The present invention can provide a method for determining authenticity that is free from the influence of noise in non-dot areas and is not limited by the type of reading device. [Brief explanation of the drawings]

[0014] [Figure 1] Schematic diagram of how the learning model is created [Figure 2] Schematic diagram of the method for determining whether a product is genuine or not, using machine learning [Figure 3] Image of an image generated for machine learning in an embodiment [Figure 4] Image of the Fourier transform for machine learning in the example [Figure 5] Image after Fourier transform for machine learning in other embodiments [Figure 6] An example of a confusion matrix [Figure 7] Schematic diagram of how the classification of the machine learning data is determined from a vector consisting of the similarity of each classification item [Figure 8] Similarity distribution [Figure 9] A schematic diagram of determining the classification of the machine learning data by setting a similarity threshold from a vector consisting of the similarity of each classification item. DETAILED DESCRIPTION OF THE INVENTION

[0015] The following description will discuss embodiments of the present invention with reference to the accompanying drawings. However, the present invention is not limited to the embodiments described below, and various other embodiments are also encompassed within the scope of the technical concept set forth in the claims.

[0016] (Learning model creation method) First, the learning model creation method of the present invention will be described. As shown in Figure 1, the learning model creation method includes a first image data acquisition step (S100) of acquiring image data of a plurality of images of genuine products and a plurality of images of non-genuine products, a first frequency image conversion step (S101) of performing frequency conversion on the image data of the genuine products and the image data of the non-genuine products to create frequency-converted image data of the genuine products and frequency-converted image data of the non-genuine products, and a learning model creation step (S102) of executing a machine learning algorithm on the frequency-converted image data of the genuine products and the frequency-converted image data of the non-genuine products, and deriving the authenticity similarity between the frequency-converted image data of the genuine products and the frequency-converted image data of the non-genuine products based on the machine learning algorithm to create a learning model (300).

[0017] (First image data acquisition step (S100)) The first image data acquisition step (S100) acquires image data of both an image of the genuine product and an image of the non-genuine product. The means for acquiring both image data is not particularly limited and may be a device having a known imaging means such as a camera or scanner. The format of the acquired image data is not particularly limited and may be, for example, JPEG, PNG, BMP, or other formats.

[0018] The authentic product in the present invention may be a printed matter or paper with elements having a predetermined regularity. Specifically, any medium having an image from which frequency-converted image data can be obtained may be used, such as a printed matter with halftone dots, lines, etc., paper formed by a wire method such as a Fourdrinier paper machine, a cylinder paper machine, a twin-wire paper machine, or a hand-made paper machine, a diffraction grating formed by demetallization, a relief hologram, or an image formed by an array of laser perforations. The aforementioned printed matter and paper are widely used for valuable printed matter, and therefore are preferred because they allow for the creation of the learning model of the present invention on many occasions. Non-authentic products include copies or replicas of the aforementioned authentic products, and the method of copying and replicating is not limited.

[0019] The reason for using image data of genuine and non-genuine products is to run the machine learning algorithm described below on both sets of image data, derive their similarities, and use them as image classification data for machine learning for each classification category of genuine and non-genuine products.

[0020] Note that the classification of images by machine learning in the present invention is a method called supervised learning, in which images and the names of the items to which the images belong (also called "labels") are used in advance as machine learning data for learning. As a result of machine learning, a value called similarity is output for each item for an unknown image. Generally, the item with the highest similarity is determined to be the item to which the input image belongs. Furthermore, to perform machine learning with higher accuracy, the more genuine and non-genuine image data there are, the better.

[0021] Since the conditions for acquiring image data of the items to be identified (described later) are random, the image data of genuine and non-genuine items are acquired by acquiring multiple image data with different resolutions or randomizing the conditions for reading genuine and non-genuine items. For example, image data is acquired by randomizing the reading position, such as the height and angle of the imaging means, and the reading illuminance. Furthermore, the acquired image data may be subjected to image processing to remove noise, blur, sharpen, etc.

[0022] The image data of both genuine and non-genuine products may be image data of the entire product, and may be image data obtained by acquiring the entire genuine and non-genuine products at once, or image data obtained by dividing the genuine and non-genuine products into images of multiple regions of a specified size and combining all of the images of the divided regions.

[0023] Furthermore, since it is preferable to have a large amount of learning data in order to improve the accuracy of machine learning, a method of duplicating multiple portions of image data from a single or multiple sets of genuine and non-genuine image data may also be used.

[0024] In this embodiment, the images are classified into two categories, genuine and non-genuine, but it is also possible to classify them based on the environment used in the discrimination method, the camera specifications, and the state of the camera and the subject (the distance between them, the reading angle, etc.). For example, by classifying images of genuine products at multiple resolutions and images of non-genuine products at multiple resolutions, highly accurate discrimination can be achieved.

[0025] (First frequency image conversion step (S101)) The first frequency image conversion process (S101) applies a known frequency conversion method such as Fourier transform or wavelet transform to the image data of the genuine and non-genuine products acquired in the first image data acquisition process (S100) to create frequency converted image data of the genuine product and frequency converted image data of the non-genuine product.

[0026] (Learning model generation process (S102)) The learning model generation process (S102) executes a machine learning algorithm on the frequency-converted image data of the genuine product and the frequency-converted image data of the non-genuine product, and based on the machine learning algorithm, derives the degree of authenticity similarity between the frequency-converted image data of the genuine product and the frequency-converted image data of the non-genuine product to generate a learning model (300).

[0027] Machine learning algorithms refer to computer algorithms used to automatically extract useful information from machine learning data by constructing known probability models (also called machine learning models, or simply learning models). Machine learning is performed using one or more learning algorithms, such as classification and regression techniques (e.g., support vector methods, trees, neural networks, etc.).

[0028] For example, a neural network consists of a combination of convolutional layers, pooling layers, dropout layers, flattening layers, and fully connected layers. The convolutional layer extracts the features of each input image using a randomly generated filter. The pooling layer outputs a reduced image by reducing the size of the output data while retaining the features of the image. The dropout layer disables irrelevant regions from the input image. The flattening layer converts the input image into a one-dimensional array and outputs it, and is used before the next fully connected layer. Then, the fully connected layer uses a function called Softmax to output a score that represents the similarity for each classification item of genuine and non-genuine products, with the closer to "1" the score, the more similar it is.

[0029] These processes are repeated a number of times indicated by the number of epochs (learning is repeated multiple times using each piece of machine learning data. This number is called the "epoch number E") to make the similarity more appropriate, thereby generating a machine learning model (300).

[0030] The above-mentioned machine learning is a method for generating a machine learning model that classifies products into two categories: genuine and non-genuine, but depending on the environment of the discrimination method, a machine learning model (300) that performs three or more classifications, such as classifying genuine products into multiple resolutions, may be generated. Other classifications possible include various classifications based on the angle of the print medium and camera, the state of image focus (so-called focus), and even brightness.

[0031] As a method for calculating the similarity, for example, SAD (Sum of Absolute Differences), which is one of the template matching methods, or SSD (Sum of Squared Differences) can be applied. Note that, unlike this embodiment in which the closer to "1" the similarity is determined to be higher, in the case of a method using SAD or SSD, the similarity is evaluated by the sum of squares of the differences in pixel values ​​of A and B, and it can be determined that the closer to "0" the similarity is higher.

[0032] (True / false determination method) Next, the authenticity discrimination method of the present invention will be described. As shown in Fig. 2, the authenticity discrimination method of the present invention comprises a second image data acquisition step (S200) of acquiring image data of at least a portion of the item to be discriminated, a second frequency image conversion step (S201) of frequency-converting the image data of the item to be discriminated to create frequency-converted image data of the item to be discriminated, a similarity deriving step (S202) of inputting the frequency-converted image data of the item to be discriminated to a learning model (300) and deriving the similarity of the frequency-converted image data, and a discrimination step (S203) of discriminating whether the item to be discriminated is authentic or non-authentic based on the result of the similarity deriving step (S202).

[0033] (Second image data acquisition step (S200)) The second image data acquisition step (S200) acquires image data of the target item from the target item. The means for acquiring the image data of the target item is not particularly limited and may be a device having a known imaging means such as a camera or scanner. The format of the acquired image data is not particularly limited and may be, for example, JPEG, PNG, or BMP.

[0034] The size of the image data to be acquired is not particularly limited and may be different from that of the first image data acquisition step (S100) described above, but is preferably the same size as the image data before frequency conversion in the first frequency image conversion step (S101). However, if the acquired image data is larger than the image data before frequency conversion, it can be converted to the image size before frequency conversion by cropping or the like. Furthermore, even if the size of the acquired image data is smaller than the image data before frequency conversion in the first frequency image conversion step (S101), an image that appears out of focus is acceptable, and in recent years, it has become possible to enlarge images while suppressing degradation using AI, so the data size of the image is not restricted.

[0035] (Second frequency image conversion step (S201)) In the second frequency image conversion step (S201), a known frequency conversion method such as Fourier transform or wavelet transform is applied to the image data of the item to be identified acquired in the second image acquisition step (S200), to create frequency-converted image data of the item to be identified. Note that the method of the second frequency image conversion step (S201) may be different from that of the first frequency image conversion step (S101), but it is preferable that they are the same in terms of accuracy of identification.

[0036] (Similarity derivation step (S202)) The similarity deriving step (S202) is a step of deriving the similarity of the frequency-converted image data of the discrimination target product using the machine learning model (300). The similarity deriving is performed using the same method as the method for creating the machine learning model (300) obtained by the learning model creation method to derive the similarity of the frequency-converted image data of the discrimination target product generated in the second frequency image conversion step (S201) and derive an approximation to the similarity of either a genuine product or a non-genuine product in the machine learning model (300), which has the advantage of not requiring a large storage area.

[0037] For example, when the number of classification items of the used machine learning model (300) is N, an N-dimensional similarity vector is output for the frequency-converted image data of the discrimination target product generated in the second frequency image conversion step (S201). Each component P n (n = 1 to N) represents the probability of being the nth (n = 1 to N) item, and P1 + P2 + ··· + P N = 1 holds.

[0038] In the case of the machine learning model (300) classified into two types of items, genuine and non-genuine, when N = 2 and the similarity vector is (P1, P2), for example, when learning is performed using a genuine image as the first component and a non-genuine image as the second component during machine learning, P1 represents the probability of being a genuine image and P2 represents the probability of being a non-genuine image.

[0039] (Discrimination step (S203)) The discrimination step (S203) is a discrimination step (S203) that discriminates whether the discrimination target product is a genuine product or a non-genuine product based on the result of the similarity derivation step (S202). Specifically, from the components of each similarity vector created by the similarity derivation step (S202), the component indicating the one with the closest similarity is determined. For example, in the case of a machine learning model where the closer the similarity is to "1", the more similar it is, it is determined to belong to the classification item with the maximum value.

[0040] More specifically, when performing true / false discrimination using a machine learning model (300) that can classify two types of images, genuine and non-genuine, and the machine learning is performed with the first classification item being "genuine" and the second classification item being "non-genuine", for the image of the discrimination target product, if P1 > P2 in the similarity vector (P1, P2) derived through the second image data acquisition step (S200), frequency image conversion step (S201), and similarity derivation step (S202), the discrimination target product is determined to be genuine; if P1 < P2, the discrimination target product is determined to be non-genuine; and if P1 = P2, it is determined that the discrimination is impossible.

[0041] Furthermore, to perform true / false discrimination more precisely, a threshold value for similarity is set, and if the maximum similarity exceeds the threshold, the classification item with the maximum threshold is determined to be the classification to which the image belongs, and if it is below the threshold, it is deemed indistinguishable and discrimination processing is performed again. This method eliminates cases where even if the maximum similarity among the N similarities is the largest, the difference with the similarities of other classification items is small, allowing for highly accurate discrimination processing.

[0042] In a method for determining authenticity using a machine learning model created with N=2, the first classification item being "genuine" and the second classification item being "non-genuine," the threshold is T (T≧0.5). After the target product has been identified using the above method, if, for example, P1>P2, a comparison is made with the threshold T, and if P1>T, the target product is determined to be authentic, and if P1≦T, it is deemed indistinguishable.

[0043] Furthermore, in the above-described embodiment, one image data item is acquired from the item to be classified, and one similarity vector is generated using the learning model (300) to perform a true / false discrimination. However, it is also possible to acquire multiple image data items from the item to be classified, generate similarity vectors for each, and make a comprehensive judgment by adding up the similarity vectors for each Nth classification item, for example, and using the maximum value as the classification result for the object.

[0044] For example, in a method for determining authenticity using a machine learning model created with N=2 and the first classification item as "genuine" and the second classification item as "non-genuine," when determining authenticity using multiple image data (M pieces of image data (M is an integer of 2 or more)) obtained from different locations on the item to be determined, the similarity vector obtained from the image data obtained from the first location on the item to be determined is (P 11 ,P 12 ), and the similarity vector obtained from the image data acquired from the first location and the Mth location different from the first location of the same discrimination target product is defined as (P M1 ,P M2 ), then P 11 +P 21 +···+P M1 >P 12 +P 22 +···+PM2 If so, it is determined to be genuine, and P 11 +P 21 +P M1 <P 12 +P 22 +···+P M2 If so, non-authentic, P 11 +P 21 +P M1 =P 12 +P 22 +···+P M2 If so, it is deemed indistinguishable.

[0045] Compared to conventional methods that focus on the characteristics of genuine products, the authenticity determination method of the present invention utilizes the characteristics of both genuine and non-genuine products, and since the characteristics become apparent by performing frequency transformation such as Fourier transform on image data of printed matter composed of halftone dots, etc., it is possible to improve accuracy by generating a learning model and performing authenticity determination on the frequency-transformed image.

[0046] (software) Next, a program for causing a computer to execute the learning model creation method and the discrimination method of the present invention will be described.

[0047] (Software for creating learning models) As the software for the learning model creation method of the present invention, known software such as Tensorflow Lite Model Maker, an open source software provided by Google (registered trademark), can be used. In addition, the use of transfer learning, which is a machine learning technique, reduces the amount of training data required and shortens the time spent on training.

[0048] (Software for determining whether something is true or false) Furthermore, the software for executing the method for determining whether a given parameter is true or false may be software that can use the machine learning model (300) generated in the learning model generation step (S102) in the machine learning model creation method. For example, if the learning model is created using the Keras neural network library, the prediction function predict is used.

[0049] Specifically, image data of the target product captured by a scanner or the like is subjected to frequency transformation processing such as Fourier transform using the programming language Python, and the transformed image data and the machine learning model are input into the prediction function predict. A similarity vector is output, and software can be obtained by programming the discrimination step (S203) using this similarity vector.

[0050] Alternatively, you can create software to distinguish between genuine and fake using Android Studio, a smartphone app development tool from Google. While we won't go into details here, to perform on-device discrimination on a smartphone with a genuine and fake app that includes a machine learning model installed, you can import a machine learning model created through transfer learning using Tensorflow Lite Model Maker, an open-source software provided by Google, into a PC with Android Studio installed, and use a library called CameraX to handle camera processing.

[0051] CameraX is capable of image analysis in real time. By using this, images can be collected in the camera's preview state, where the subject is always displayed, frequency converted within the device, and used as the input image to derive a similarity vector. Image classification, i.e., true / false determination, can then be performed in real time from the similarity vector.

[0052] Next, examples of the present invention will be described, but the embodiments of the present invention are not limited to these examples.

[0053] (Example) In this embodiment, it is assumed that the machine learning model is recognized by the smartphone application development software AndroidStudio, and the smartphone application is used to determine whether the model is true or false.

[0054] First, a commercially available halftone print was used as the authentic sample, and a non-authentic sample was a copy of the authentic sample printed in map mode on a digital full-color copier (RICOH Pro C7100S).

[0055] Next, as the first image data acquisition step (S100), the above-mentioned digital full-color copier was used to acquire genuine and non-genuine image data of 600 dpi and 200 dpi using a scanner at resolutions of 600 dpi and 200 dpi for each of the genuine and non-genuine products.

[0056] Furthermore, the 600 dpi genuine product image data, 200 dpi genuine product image data, 600 dpi non-genuine product image data, and 200 dpi non-genuine product image data were each randomly rotated and resized to a size of 64 pixels x 64 pixels, obtaining 10,000 pieces of each image data, for a total of 40,000 pieces of image data.

[0057] FIG. 3 shows examples of four types of resized image data: 600 dpi genuine image data (12), 200 dpi genuine image data (22), 600 dpi non-genuine image data (32), and 200 dpi non-genuine image data (42).

[0058] Next, as the first frequency image transformation step (S101), a two-dimensional Fourier transform method was applied to the 40,000 pieces of image data acquired in the first image data acquisition step (S100) to create two-dimensional Fourier transform image data for genuine and non-genuine products.

[0059] Next, as an example, Figure 4 shows two-dimensional Fourier transform image data (13) of 600 dpi genuine product image data (12), two-dimensional Fourier transform image data (23) of 200 dpi genuine product image data (22), two-dimensional Fourier transform image data (33) of 600 dpi non-genuine product image data (32), and two-dimensional Fourier transform image data (43) of 200 dpi non-genuine product image data (42).

[0060] Although this embodiment describes halftone prints, it is also possible to create two-dimensional Fourier transform image data (13-1) of 600 dpi genuine image data (12-1) of a print with lines formed on it as shown in Figure 5, or two-dimensional Fourier transform image data (13-2) of 600 dpi genuine image data (12-2) of paper formed by a papermaking process such as a wire method.

[0061] Next, in the learning model generation process (S102), a machine learning algorithm was applied to the 40,000 pieces of image data created in the first frequency image conversion process (S101), and a learning model (300) that derives similarity was generated based on the machine learning algorithm.

[0062] In this example, a neural network library written in Python called Keras, which is a machine learning algorithm that is publicly available in literature, on the Internet, etc., was used.

[0063] Specifically, machine learning was performed for 10 epochs, with one set consisting of a convolutional layer, a pooling layer, a dropout layer, a flattening layer, and a fully connected layer.

[0064] FIG. 6 shows a confusion matrix in which the similarities of the two-dimensional Fourier transform image data (13) of the 600 dpi genuine product image data (12), the two-dimensional Fourier transform image data (23) of the 200 dpi genuine product image data (22), the two-dimensional Fourier transform image data (33) of the 600 dpi non-genuine product image data (32), and the two-dimensional Fourier transform image data (43) of the 200 dpi non-genuine product image data (42) were calculated and classified into the ones with the greatest similarity, as an evaluation of the performance of the machine learning model that generated the Fourier transform images in this example.

[0065] The sum of the numbers on each row is "0" on the vertical axis, which represents the two-dimensional Fourier transform image data (13) of the 600 dpi genuine product image data, "1" on the vertical axis, which represents the two-dimensional Fourier transform image data (23) of the 200 dpi genuine product image data (22), "2" on the vertical axis, which represents the two-dimensional Fourier transform image data (33) of the 600 dpi non-genuine product image data, and "3" on the vertical axis, which represents the two-dimensional Fourier transform image data (43) of the 200 dpi non-genuine product image data (42), totaling 10,000 pieces of training data obtained by two-dimensional Fourier transform for each classification item.

[0066] The estimation results for each column show the following estimation results: "0" on the vertical axis indicates the two-dimensional Fourier transform image data (13) of genuine product image data at 600 dpi; "1" on the vertical axis indicates the two-dimensional Fourier transform image data (23) of genuine product image data (22) at 200 dpi; "2" on the vertical axis indicates the two-dimensional Fourier transform image data (33) of non-genuine product image data (32) at 600 dpi; and "3" on the vertical axis indicates the two-dimensional Fourier transform image data (43) of non-genuine product image data (42) at 200 dpi.

[0067] The vertical axis indicates "0" for 10,000 two-dimensional Fourier transform image data (13) of 600 dpi genuine product image data (12), while the horizontal axis indicates "0" for 9,834 images that were guessed to be two-dimensional Fourier transform image data (23) of 600 dpi genuine product image data (12). The horizontal axis indicates "1" for 0 images that were guessed to be two-dimensional Fourier transform image data (23) of 200 dpi genuine product image data (22). The horizontal axis indicates "2" for 164 images that were guessed to be two-dimensional Fourier transform image data (33) of 600 dpi non-genuine product image data (32), and the horizontal axis indicates "2" for 2 images that were guessed to be two-dimensional Fourier transform image data (43) of 200 dpi non-genuine product image data (42).

[0068] (True / false determination method) Using the machine learning model (300) obtained by the above-mentioned learning model creation method, authenticity was determined by classifying the images obtained from the product to be identified into one of two-dimensional Fourier transform image data (13) of genuine product image data (12) at 600 dpi (where "0" on the vertical axis) or two-dimensional Fourier transform image data (23) of genuine product image data (22) at 200 dpi (where "1" on the vertical axis) or two-dimensional Fourier transform image data (33) of non-genuine product image data at 600 dpi (where "2" on the vertical axis) or two-dimensional Fourier transform image data (43) of non-genuine product image data (42) at 200 dpi (where "3" on the vertical axis)).

[0069] First, in the second image data acquisition process (S200), image data of the item to be distinguished was acquired from the item to be distinguished using the same means as in the first image data acquisition process (S100), and a single 64 pixel x 64 pixel image was cut out from the acquired image to create image data of the item to be distinguished.In the second frequency image conversion process (S201), a two-dimensional Fourier transform method was applied in the same way as in the first frequency image conversion process (S101), to create frequency-converted image data of the item to be distinguished.

[0070] Next, using the created machine learning model (300), a four-dimensional vector representing the similarity for each classification of the frequency-converted image data was generated in a similarity derivation step (S202) as shown in FIG.

[0071] Next, in the discrimination step (S203), the item with the maximum value in the four-dimensional vector was determined to be the item to which the frequency-converted image data of the discrimination target product belonged. As shown in Fig. 7, since the similarity between the 600 dpi genuine product image data (12) and the two-dimensional Fourier-transformed image data (13) was the highest at 0.85, the two-dimensional Fourier-transformed image of the discrimination target product was classified as the two-dimensional Fourier-transformed image data (13) of the 600 dpi genuine product image data (12), and the target product was determined to be genuine.

[0072] Next, an example will be described in which the accuracy of discrimination can be improved by setting a threshold value for the similarity in the discrimination step (S203).

[0073] Figure 7 shows the similarity vectors generated for each learning data set using the generated learning model, and the data are classified into the category with the highest value. Figure 8 shows the distribution of the maximum similarity values ​​in a histogram.

[0074] In the distribution of similarities shown in Figure 8, the distribution of similarities for a total of 9,834 images where the two-dimensional Fourier transform image data (13) of the 600 dpi genuine product image data (12) was judged to be the 600 dpi genuine product image data (12) is shown as "0 → 0." For example, this indicates that there were 8,869 images with similarities in the range of 0.9 to 1.

[0075] On the other hand, the distribution of 78 images that were determined to be genuine 600 dpi from the two-dimensional Fourier transform image data (33) of 600 dpi non-genuine image data (32) is shown in "2 → 0", with 46 images in the range of 0.5 to 0.6, and it can be seen that the distribution is concentrated at lower values ​​compared to the case of "0 → 0".

[0076] Therefore, as shown in FIG. 9, a more precise discrimination is possible by determining that a similarity is 0.9 or more and determining that a similarity is indistinguishable if it is less than 0.9.

[0077] (software) Next, an embodiment of a program for causing a computer to execute the learning model creation method and the discrimination method of the present invention will be described.

[0078] (Software for creating learning models) As the software for the learning model creation method of the present invention, known software such as Tensorflow Lite Model Maker, an open source software provided by Google, which utilizes transfer learning, can be used. In addition, the use of transfer learning, a machine learning technique, reduces the amount of training data required and shortens the time spent on training.

[0079] (Software for determining whether something is true or false) Furthermore, the software for executing the method for determining whether a given parameter is true or false may be software that can use the machine learning model (300) generated in the learning model generation step (S102) in the learning model creation method. For example, if the machine learning model is created using the Keras neural network library, the prediction function predict is used.

[0080] Specifically, the image data of the target product captured by a scanner or the like is subjected to frequency transformation processing such as Fourier transform using the programming language Python, and the transformed image data and the machine learning model are input into the aforementioned prediction function predict. A similarity vector is output as the output, and software can be obtained by programming the discrimination step (S203) using this similarity vector.

[0081] Alternatively, you can create software to distinguish between genuine and fake using Android Studio, a smartphone app development tool from Google. While we won't go into details here, to perform on-device discrimination on a smartphone with a genuine and fake app that includes a machine learning model installed, you can import a machine learning model created through transfer learning using Tensorflow Lite Model Maker, an open-source software provided by Google, into a PC with Android Studio installed, and use a library called CameraX to handle camera processing.

[0082] CameraX is capable of real-time image analysis, which allows images to be collected in the camera's preview state, where the subject is always displayed, and frequency converted within the device. This is then used as the input image to derive a similarity vector, which can then be used to classify the image, i.e., to determine whether it is true or false, in real time.

Claims

1. a first image data acquisition step of acquiring image data of a plurality of genuine product images and a plurality of non-genuine product images; a first frequency image conversion step of performing frequency conversion on the plurality of image data of genuine products and the plurality of image data of non-genuine products to generate frequency converted image data of genuine products and frequency converted image data of non-genuine products; A method for creating a learning model for authenticity discrimination, comprising a learning model generation step of executing a machine learning algorithm on the frequency-converted image data of the genuine product and the frequency-converted image data of the non-genuine product, and deriving the similarity of authenticity between the frequency-converted image data of the genuine product and the frequency-converted image data of the non-genuine product based on the machine learning algorithm to generate a learning model.

2. 2. The method for creating a learning model for authenticity discrimination according to claim 1, wherein the authentic product is a printed matter or paper on which elements having a predetermined regularity are applied.

3. A method for determining whether a learning model is true or false, using the learning model created by the learning model creation method according to claim 1 or 2, a second image data acquisition step of acquiring image data of at least a portion of the item to be identified; a second frequency image conversion step of frequency converting the image data of the object to be identified to generate frequency converted image data of the object to be identified; a similarity deriving step of inputting the frequency-converted image data of the discrimination target product into the learning model and deriving a similarity of the frequency-converted image data; The authenticity determination method further comprises a determination step of determining whether the item to be determined is genuine or non-genuine based on the result of the similarity deriving step.

4. 3. Software for causing a computer to execute the learning model creation method according to claim 1 or 2.

5. 4. Software for causing a computer to execute the method for determining authenticity according to claim 3.

Images