AI-based integrated art authenticity determination system

Abstract

The present invention relates to an art authentication system utilizing artificial intelligence technology, and more specifically, to a processor-based system that integrates and analyzes the visual image, material composition, and literature history data of an artwork to calculate a quantitative score regarding its authenticity.

Description

Technology Field

[0001] The present invention relates to an art authentication system utilizing artificial intelligence technology, and more specifically, to a processor-based system that integrates and analyzes the visual image, material composition, and literature history data of an artwork to calculate a quantitative score regarding its authenticity. Background Technology

[0002] Art appraisal has traditionally consisted of visual appraisal, which relies on the subjective eye of experts, and scientific appraisal, which utilizes analytical equipment. However, conventional visual appraisal lacks objectivity as results can vary depending on the appraiser's skill level and subjectivity, and there are limitations in that the appraisal results are provided in vague expressions such as "looks like a genuine piece" or "doubtful."

[0003] Furthermore, the scientific authentication methods performed by existing authentication firms tend to be skewed toward specific analytical techniques, such as Raman spectroscopy or the analysis of specific pigments. This single-method analysis has a technical limitation in that it is difficult to accurately determine authenticity if a forger meticulously mimics only specific analytical indicators.

[0004] Furthermore, there is a current lack of quantitative indicators and systems capable of organically integrating data of diverse natures—such as visual characteristics, chemical composition, and ownership history—to present the probability of authenticity as an objective figure. Consequently, securing reliable, standardized grounds for appraisal poses significant challenges during art transactions, insurance applications, and legal disputes. Prior art literature

[0005] Korean Patent Publication 10-2009-0032637 A (2009.04.01) Korean Patent Publication 10-2024-0063537 A (2024.05.10) Korean Patent Publication 10-2024-0010780 A (2024.01.25) The problem to be solved

[0006] Conventional art appraisal is skewed toward visual appraisal relying on the subjective eye of appraisers, raising concerns about a lack of objectivity. Furthermore, relying solely on specific scientific analysis methods, such as Raman spectroscopy, limits the ability to identify forgeries that have merely mimicked specific analysis items.

[0007] Furthermore, as appraisal results are provided using ambiguous expressions such as "appears authentic" or "doubtful," there is a challenge in utilizing them as a basis for clear judgment in art transactions, insurance subscriptions, and legal disputes.

[0008] In particular, in the absence of an actual operating AI model, there is a need to derive quantitative indicators by systematically designing logical causal relationships between image data, scientific component values, and literature history.

[0009] Accordingly, the present invention aims to provide a system that calculates the probability of authenticity as an objective numerical value by having a processor integrate and learn the visual characteristics, material composition, and historical data of an artwork to solve these problems. means of solving the problem

[0010] One embodiment of the present invention proposes an electronic device for determining the authenticity of a work of art, comprising: a memory for storing data regarding the work of art; and a processor for executing instructions stored in the memory, wherein the processor acquires image data, material component analysis data, and literature history data of the work of art; individually analyzes the acquired data through three independent artificial intelligence engines that each perform image analysis, material analysis, and history analysis; and calculates an Integrated Authenticity Assessment Index (IAI Score) that quantifies the probability of the work of art being authentic as a numerical value between 0 and 100 by reflecting patterns in which the analysis evidence derived from the three independent artificial intelligence engines reinforces or conflicts with each other into an analysis model.

[0011] The processor may be characterized by performing preprocessing on the image data including noise removal, color normalization, and illumination fix; and dividing the preprocessed image data into patches of 128 to 512 pixels (px) to set areas for pen and crack analysis.

[0012] The artificial intelligence engine performing the above image analysis may be characterized by including: a stroke analysis module (StrokeNet) that verifies whether the artist's unique stroke grammar matches based on the flow, curvature, and pressure map of brush strokes; a texture analysis module (TextureNet) that detects periodicity and unnatural uniformity of surface textures using deep features based on LBP (Local Binary Patterns) and CNN (Convolutional Neural Network); and a crack analysis module (CrackNet) that models the crack structure of paint layers and determines whether the cracks are artificial using U-Net and GNN (Graph Neural Network).

[0013] The artificial intelligence engine performing the above material analysis may be characterized by being configured to receive data values ​​output from at least one of X-ray fluorescence analysis (XRF), Raman spectroscopy (Raman), and infrared spectroscopy (FTIR), and to detect indirect evidence of forgery by comparing and analyzing the elemental composition, molecular structure, characteristics of the binder, and degree of aging of the pigment with reference data.

[0014] The artificial intelligence engine performing the above-mentioned history analysis may be characterized by being configured to digitize the above-mentioned document history data through OCR (Optical Character Recognition) technology and to derive indicators of Document Integrity, which compares the font and format of the document; Event Consistency, which analyzes cross-verification with an external database; and Historical Validity, which analyzes the consistency with the author's date.

[0015] The processor may be characterized by being configured to reset the integrated authenticity assessment index by applying a preset penalty logic to the conflicting patterns when pigment components inconsistent with the production date are detected in the material analysis results or traces of document tampering are detected in the history analysis results, even if the image analysis results support authenticity. Effects of the invention

[0016] According to the present invention, the objectivity and precision of the appraisal can be maximized by integrating the analysis results derived from three independent analysis modules, namely image AI, material AI, and history AI, to calculate the final authenticity probability.

[0017] By providing a quantified IAI Score between 0 and 100 instead of the ambiguous expressions of traditional appraisal methods, it has the effect of establishing standardized criteria for art trading and insurance subscription and securing clarity as a legal basis.

[0018] In addition, by cross-verifying visual verification through brushstroke, crack, and texture analysis, scientific compositional verification through XRF, Raman, and FTIR analysis, and OCR-based literature history analysis, sophisticated forgeries that bypass only specific analysis methods can be effectively blocked.

[0019] Consequently, by comprehensively reflecting patterns where evidence from the three domains reinforces or conflicts with one another, the reliability of the emotional results can be dramatically improved. Brief explanation of the drawing

[0020] Other aspects, features, and benefits of specific preferred embodiments of the present invention, as described above, will become more apparent from the following description in conjunction with the accompanying drawings. FIG. 1 is a drawing showing a system according to one embodiment of the present invention. FIG. 2 is a block diagram showing a system according to one embodiment of the present invention. FIG. 3 is a diagram illustrating an image analysis engine (Image AI) according to an embodiment of the present invention. FIG. 4 is a diagram illustrating a material component analysis engine (Material AI) according to one embodiment of the present invention. FIG. 5 is a diagram illustrating a literature history analysis engine (Provenance AI) according to one embodiment of the present invention. FIG. 6 is an AI-based quantitative authenticity appraisal index (IAI Score) according to an embodiment of the present invention. TM It is a drawing to explain ). It should be noted that in the drawings above, similar reference numbers are used to illustrate identical or similar elements, features, and structures. Specific details for implementing the invention

[0021] Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. To clarify the gist of the invention, irrelevant technologies are omitted, and components within the drawings may be depicted schematically. The present invention is defined only by the scope of the claims and is not limited to the disclosed embodiments, and like reference numerals refer to like components.

[0022] Each block of the flowchart can be implemented as a computer program instruction executed by a processor, and such instructions are loaded into a general-purpose or special-purpose computer processor to create a means for performing a specific function. Each block refers to a module or part of code that executes a specific logical function, and unlike the illustrated order, it may be executed substantially simultaneously or in reverse order.

[0023] The term "part" in this specification refers to a hardware or software component configured on an addressable storage medium or implemented to drive one or more processors. The functions provided by such components and "parts" may be applied to other systems and services having a similar technical background without departing from the scope of the present invention.

[0024] FIG. 1 is a drawing showing a system according to an embodiment of the present invention, wherein the system of the present invention comprises a server (110) and a terminal (120). Meanwhile, in the present invention, the server (110) may be referred to as an electronic device.

[0025] The system of the present invention is the world's first three-stage integrated AI appraisal platform, which organically combines image analysis AI (Image AI), material analysis AI (Material AI), and history analysis AI (Provenance AI) technologies to calculate and provide an AI-based quantitative authenticity appraisal indicator called the IAI Score. Referring to FIG. 1, a server (110) according to one embodiment of the present invention generates authenticity determination information, which is information regarding AI-based integrated authenticity determination of artworks, and provides it to a terminal (120), and the user of the terminal (120) can check the authenticity determination information through the output device of the terminal (120).

[0026] The server (110) integrates and learns the results of three different areas of analysis—image, which is a visual characteristic of the artwork; material, which is a scientific component analysis; and history based on literature and data—through artificial intelligence, and provides a numerical value of the probability of authenticity on a score scale ranging from 0 to 100 points. Unlike existing appraisal companies that insisted on specific methods such as Raman spectroscopy or expert subjective visual appraisal, this ensures objectivity by integrating all analysis data, which is a key differentiating factor. At this time, the control module (111, 121) of the system or at least one processor performs the role of implementing or controlling the image analysis engine, material component analysis engine, literature history analysis engine, and the logic for calculating an integrated authenticity determination indicator.

[0027] More specifically, the Image AI engine acquires the surface image of the artwork as high-resolution data and performs analysis by dividing it into three modules: Stroke, Texture, and Craquelure. The Stroke Net analysis module databases the artist's unique brushstroke style and signature patterns to detect forgery detection signals through curvature, entropy, and pressure estimation, while the Texture Net analysis module analyzes unnatural surface uniformity or printing periodicity using LBP, Gabor filters, and CNNs. Additionally, the Craquelure analysis module analyzes paint layer cracking patterns using graph statistics and morphology to determine the consistency between the presence of artificial cracks and the claimed date of the artwork.

[0028] The Material AI engine automates and quantifies precision scientific analysis methods, such as X-ray fluorescence analysis (XRF), Raman spectroscopy, and Fourier Transistor (FTIR), using artificial intelligence. The processor receives data values ​​related to elemental composition, molecular structure, binder, and the degree of maturation from each instrument and compares them with reference data to determine whether pigment patterns match by era or artist, and to detect indirect evidence of forgery, such as prohibited pigments or traces of printing.

[0029] The Provenance AI engine digitizes and analyzes text inscribed on artworks, such as records, signatures, certificates of authenticity, and auction and exhibition records, using OCR (Optical Character Recognition) technology. During the analysis process, it calculates the reliability of the literature by comprehensively assessing the consistency of the entire document, the feasibility of the claimed history, its consistency with the era and place names, and traces of image-based editing or synthesis. Additionally, it verifies the compatibility of the records by performing cross-verification with external online databases.

[0030] The final derived IAI Score is calculated by integrating evidence generated by the three independent artificial intelligence engines mentioned above, and forms a reliability-based quantitative indicator by reflecting reinforcement or conflict patterns between complex data rather than single evidence. For example, the calculated score provides interpretation criteria for intervals ranging from a very high probability of authenticity of 90 points or more to a very low probability of authenticity of less than 30 points, thereby moving away from the ambiguous expressions of traditional appraisals and presenting standardized appraisal results that can be used as a basis for transactions, insurance, and legal matters.

[0032] FIG. 2 is a drawing showing a system according to one embodiment of the present invention.

[0033] Referring to FIG. 2, the system (100) of the present invention includes a server (110) and a terminal (120). The server (110) includes a first control module (111), a first communication module (112), a first input module (113), and a first output module (114), and the terminal (120) may include a second control module (121), a second communication module (122), and a second input module (123).

[0034] A control module (111, 121) directly or indirectly controls a server (110) or a terminal (120) to implement operations, steps, and processes according to an embodiment of the present invention. In particular, the control module (111, 121) includes at least one processor, and the processor includes at least one of a central processing unit (CPU) and a graphics processing device (GPU) to perform overall computation and command execution of the system. The control module loads various programs and data stored in a storage module to be described later to process analysis logic or control lower hardware configurations.

[0035] The communication module (112, 122) transmits and receives data and information between the server (110) and the terminal (120) and can communicate with external devices through various networks including cellular, short-range wireless communication (Bluetooth, WiFi, etc.) or wired LAN. This communication configuration can be integrated into a single chip or implemented as multiple separate components.

[0036] The input module (113, 123) receives commands and data from a user or administrator and includes a touch recognition display, a touchpad, a button, a voice recognition sensor, a mouse or keyboard, etc. Here, touch recognition is configured to detect physical contact of the user through a pressure-sensitive or capacitive method.

[0037] The output module (114, 124) visually or audibly displays voice, information, data, images, and various objects generated by the control module (111, 121) or obtained through the communication module, and includes a display, speaker, and LED lamp, etc.

[0038] Additionally, the system includes a storage module that stores a basic program for operation, an application program, configuration information, and the user's personal information (ID, password, etc.). The storage module may include at least one of various types of storage media such as flash memory, a hard disk, RAM, and ROM, and the control module (111, 121) performs the authenticity determination logic of the present invention by utilizing the resources stored in the storage module.

[0040] FIG. 3 is a diagram illustrating an image analysis engine (Image AI) according to an embodiment of the present invention.

[0041] The aforementioned control module (111, 121) or at least one processor can implement and / or control an image analysis engine (Image AI).

[0042] The image analysis engine acquires surface images of the artwork as high-resolution data, and the processor performs preprocessing on the acquired image data, including noise removal, color normalization, and illumination fix. The preprocessed data undergoes a patch tiling step in which it is divided into patches of 128, 256, or 512 pixels (px), and the processor executes stroke / crack zoning for stroke and crack analysis based on this. The image analysis engine is further divided into three modules—stroke, texture, and crack—to perform complementary analyses, and the processor integrates the analysis results of each module to determine authenticity.

[0043] StrokeNet, a handwriting analysis module, analyzes the artist's unique brush stroke flow, curvature, and pressure map, and compares them against the artist's signature pattern in a database. The processor applies entropy analysis, curvature pattern analysis, and pressure estimation techniques to verify whether the input image matches the artist's handwriting grammar, and if artificial repeatability appears in the stroke, or if a signal is detected that has significantly low information complexity and is inconsistent with the artist's unique curvature pattern, it recognizes this as a forgery detection signal.

[0044] TextureNet, a texture analysis module, assists in dating by analyzing the surface texture of the artwork and the layering of the canvas or paper. The processor identifies the surface texture of the artwork using deep feature extraction techniques based on Local Binary Patterns (LBP), Gabor filters, and Convolutional Neural Networks (CNN). Through this, it determines authenticity by detecting unnatural uniformities that are difficult to find in handwork or periodic patterns caused by specific printing techniques.

[0045] CrackNet, a crack analysis module, determines the aging state of an artwork by analyzing the crack structure of the paint layers. The processor applies U-Net, GNN (Graph Neural Network), and graph statistical techniques to model crack graphs and morphological features, thereby estimating the production date. In particular, it determines whether the cracks are artificial caused by chemical treatment or deterioration due to natural aging, and generates a negative signal if there is a discrepancy between the artwork's claimed date and the analyzed aging pattern.

[0046] In this way, the feature values ​​extracted from each module are finally integrated through the Artist Matching and Style Embedding processes in the Feature Fusion stage. Based on the integrated image feature data, the processor calculates an Image AI Score quantified as a numerical value between 0 and 100 and provides it to the user.

[0048] FIG. 4 is a diagram illustrating a material component analysis engine (Material AI) according to one embodiment of the present invention.

[0049] The aforementioned control module (111, 121) or at least one processor can implement and / or control the material component analysis engine (Material AI).

[0050] The above material component analysis engine is an authentication engine that determines authenticity based on the material spectrum of an artwork, and performs technology that automates and quantifies traditional scientific analysis methods—X-ray fluorescence analysis (XRF), Raman spectroscopy (Raman), and Fourier transform infrared (FTIR) spectroscopy—through a processor.

[0051] The processor receives raw data related to elements, molecules, media, and age obtained through expensive precision analysis equipment and performs comparative analysis with reference data. Specifically, X-ray fluorescence analysis (XRF) precisely analyzes the elemental composition of the artwork in a non-destructive manner to verify whether it matches the unique pigment patterns of each era or artist, serving as a key indicator for capturing indirect evidence of forgery. Raman spectroscopy determines the legitimacy of the technique by clearly distinguishing between natural and synthetic pigments through molecular and crystal structure analysis, and enables precise authentication at the molecular level in conjunction with Fourier Transistor Infrared (FTIR).

[0052] In addition, the processor identifies the characteristics of the binder and medium by analyzing organic and inorganic components through Fourier Transistor (FTIR) spectroscopy, thereby estimating the technique and age of the artwork. In particular, FTIR plays a role in detecting whether an artificial aging process has been applied by analyzing the presence of natural aging. These multifaceted analysis items are comprehensively reviewed in terms of elemental composition, molecular structure, binder and medium, and degree of aging, and the processor scans and quantifies in real time decisive evidence of forgery, such as the use of prohibited pigments, signals of age discrepancies, or printing and printing traces.

[0053] Finally, the material composition analysis engine does not rely on the analysis results of a single piece of equipment, but integrates data derived from XRF, Raman, and FTIR to determine the suitability with the pigment pattern of the era and the artist, thereby dramatically improving the objectivity of the appraisal.

[0054] In this way, the processor includes an interface that receives digital raw data directly generated from analysis equipment (XRF, Raman, FTIR, etc.) through a communication module, and can perform an automated identification process in which artificial intelligence directly compares the peak value and waveform of the raw data with reference spectrum data to quantitatively identify pigment components without subjective interpretation by an expert.

[0056] FIG. 5 is a diagram illustrating a literature history analysis engine (Provenance AI) according to one embodiment of the present invention.

[0057] The aforementioned control module (111, 121) or at least one processor can implement and / or control a literature history analysis engine (Provenance AI), and the literature history analysis engine processes text information contained in the art collection history, auction records, exhibition records, related documents and certificates, etc., by digitizing it through OCR (Optical Character Recognition) technology.

[0058] The processor compares digitized document data with data in an internal database and an external network, and performs the function of automatically detecting forgery by precisely analyzing consistency, chronology, style, formatting, and paper patterns between internal and external data. Specifically, the processor calculates a Document Integrity Score to evaluate the consistency of the entire document, and verifies the physical legitimacy of the document by precisely comparing the fonts, formats, structures, and metadata used in the document.

[0059] Furthermore, the processor generates an Event Consistency Score to examine whether specific historical events claimed in the literature are actually plausible; this is accomplished by verifying the authenticity of the events through cross-validation with an external database. In addition, during the Historical Validity Score calculation stage, the processor analyzes the consistency between the times, place names, author's life cycle, and chronology recorded in the literature to detect points where chronological discrepancies occur.

[0060] The processor derives a Forgery Trace Score through image-based technical analysis and scientifically determines the presence of traces of editing or synthesis by analyzing noise patterns within the document. Finally, by combining all the aforementioned analysis signals to calculate a Provenance Reliability Score, it comprehensively evaluates the reliability of the entire documentation associated with the artwork, utilizing this as foundational data for authenticity verification.

[0061] As such, the literature history analysis engine of the present invention integrates physical image analysis and data-based logical analysis, thereby overcoming the limitations of simple visual appraisal and enabling objective verification of the history of artworks.

[0062] In addition, the processor includes a web crawler that collects the ownership and transaction history of the artwork in real time from online auction sites, museum exhibition archives, and public data portals, and determines whether the history is falsified by verifying the temporal and spatial compatibility between the collected online data and the submitted literature data.

[0064] FIG. 6 is an AI-based quantitative authenticity appraisal index (IAI Score) according to an embodiment of the present invention. TM It is a drawing to explain ).

[0065] The aforementioned control module (111, 121) or at least one processor may calculate and / or reset a quantitative authenticity appraisal index (IAI Score). A processor according to an embodiment of the present invention is configured to generate an integrated AI-based appraisal index that quantifies the authenticity probability of an artwork to a numerical value (scale) between 0 and 100. This aims to establish a clear index as a basis for art transactions, insurance calculations, and legal grounds by objectively standardizing the appraisal results, thereby moving away from subjective and ambiguous expressions such as "looks like authentic" or "doubtful" used in traditional appraisal methods.

[0066] The processor does not rely on a single piece of evidence, but rather calculates a final score by organically integrating evidence derived from three independent artificial intelligence engines. Specifically, the processor analyzes the naturalness of the external appearance through Image AI, verifies the consistency of the manufacturing dates of materials and media through Material AI, and evaluates the reliability of literature and historical data through Provenance AI. In this process, the processor does not stop at simply calculating the average value of each indicator, but also incorporates patterns in which detailed evidence reinforces or conflicts with one another into the analysis model to derive a final reliability-based appraisal result.

[0067] Looking at the detailed factors considered by the processor for calculating the above score, the image analysis stage first reflects the consistency of curvature and pressure patterns through stroke analysis (StrokeNet), the detection of printing periodicity and unnatural uniformity through texture analysis (TextureNet), and the presence of artificial cracks and age discrepancies through crack analysis (CrackNet).

[0068] StrokeNet, a core module of the Image AI, quantifies the physical characteristics of brush strokes from the surface image of an artwork and compares them with the artist's unique habits. Specifically, the processor calculates the entropy and curvature of the brush's trajectory and estimates the pressure map applied by the brush to the canvas from pixel data, comparing it with the artist's unique brushwork style. In particular, the StrokeNet automatically detects signature areas within the input image and compares them with the artist's signature database (Signature DB) stored in memory. At this time, the processor uses artificial intelligence to probabilistically calculate whether the stroke order, slant, and subtle changes in pressure of the signature conform to the artist's unique 'brushstroke grammar.' If mechanical repeatability appears in the brushstroke or a pattern with significantly low information complexity is detected, this is recognized as a forgery detection signal, and the image analysis score is lowered.

[0069] TextureNet analyzes the surface texture of the artwork by extracting Local Binary Patterns (LBP), Gabor filters, and CNN-based deep features from high-resolution image patches. Through this, the processor detects periodic halftone patterns or unnatural uniformity characteristic of printing methods that are difficult to occur in manual work, and analyzes the layering of the canvas or paper to utilize as an auxiliary indicator for dating. CrackNet statistically models the cracking structure of the paint layers by applying U-Net and Graph Neural Network (GNN) technologies. Based on the geometric shape and morphological characteristics of the extracted crack graph, the processor determines whether natural aging has occurred; if artificial crack signals caused by chemical treatment are detected or if there is a discrepancy between the claimed date of the artwork and the degree of physical deterioration, it is included as a factor in suspecting forgery. In this case, the image analysis score (S_{Image}) can be calculated by the following Equation 1.

[0070] [Mathematical Formula 1]

[0071]

[0072] S_{Image} represents the integrated image authenticity score finally derived through the image analysis engine (Image AI).

[0073] α_1, α_2, and α_3 represent dynamic weights individually assigned to the stroke analysis module (StrokeNet), texture analysis module (TextureNet), and crack analysis module (CrackNet), respectively. The weights are basically set to satisfy the normalization condition α_1 + α_2 + α_3 = 1, but the processor is configured to adaptively redistribute the weights when it is determined that a specific analysis module has high discriminative power depending on the production technique of the artwork to be analyzed (e.g., oil painting, ink painting, printmaking, etc.) (e.g., in the case of ink painting where the uniqueness of the stroke is emphasized, the weight of α_1 is set relatively high).

[0074] S_{Stroke}: A stroke authenticity score representing the degree of agreement between the artist's unique brush trajectory and signature pattern calculated through the stroke analysis module (StrokeNet).

[0075] S_{Texture}: This is a texture authenticity score that quantifies the detection results of surface texture periodicity and printing techniques generated by the texture analysis module (TextureNet).

[0076] S_{Crack}: This is an aging consistency score that quantifies the degree of consistency between the crack structure of the paint layer calculated through the crack analysis module (CrackNet) and the natural aging state.

[0077] The process of calculating the final image integration authenticity score (S_{Image}) in an image analysis engine (Image AI) according to one embodiment of the present invention is as follows. The image analysis engine performs preprocessing, including noise removal, color normalization, and lighting correction, on the acquired image data of the artwork, and divides it into patches of specific pixels to perform analysis for each individual module. At this time, each parameter of the following mathematical formula 1, which serves as the basis for deriving the final score, is quantified through the computational procedure of an independent artificial intelligence engine.

[0078] First, the stroke authenticity score (S_{Stroke}) is a value calculated by the stroke analysis module (StrokeNet), which primarily analyzes the artist's unique brush stroke flow, curvature, and pressure map. The processor manages the artist's unique brushstroke style and signature pattern by creating a database, performs entropy analysis and curvature pattern analysis of the brush trajectory from the input image, and verifies whether it matches the artist's stroke grammar by applying pixel data-based pressure estimation technology. In particular, the StrokeNet automatically detects signature areas within the input image and compares them with the artist's signature database by period; if mechanical repeatability appears in the stroke or a signal with significantly low information complexity is detected, it recognizes this as a forgery detection signal and lowers the score.

[0079] Meanwhile, the texture authenticity score (S_{Texture}) is derived through a texture analysis module (TextureNet), which plays a role in assisting in dating by analyzing the texture of the artwork's surface and the layering of the canvas or paper. The processor identifies fine textures on the artwork's surface by utilizing deep feature extraction techniques based on LBP (Local Binary Patterns) and CNN (Convolutional Neural Network), and thereby determines authenticity by detecting unnatural uniformity that is difficult to achieve through manual work or periodic patterns caused by specific printing techniques.

[0080] In addition, the aging consistency score (S_{Crack}) is a value calculated by the crack analysis module (CrackNet), which determines the actual aging state of the artwork by analyzing the crack structure of the paint layer. The processor applies U-Net and GNN (Graph Neural Network) to model the crack graph and morphological features, thereby identifying whether the crack is artificial caused by chemical treatment or aging due to the natural passage of time. If the structural similarity between the artwork's claimed production date and the analyzed aging pattern is below a preset threshold, the processor recognizes this as a negative signal and reflects it in the score calculation model.

[0081] In the above mathematical formula 1, the dynamic weights (α_1, α_2, α_3) applied to individual scores are adaptively redistributed by the processor according to the production technique of the artwork to be analyzed. Basically, the sum of the weights is set to 1, but for groups of works such as ink paintings where the uniqueness of the brushwork is the core of the appraisal, the weight (α_1) of the brushwork analysis module is set relatively high, and for groups of works such as oil paintings where surface texture and physical aging are important, the weights of texture and crack analysis are adjusted upward to maximize the precision of the identification.

[0082] As an actual operation example for the score calculation logic above, we assume a case where image analysis is performed on a traditional ink painting where the artist's unique brushwork style is important. The processor dynamically assigns weights α_1=0.6, α_2=0.2, and α_3=0.2 considering the characteristics of the artwork. At this time, if the signature and brushstroke match score by the stroke analysis module (StrokeNet) is 92 points (S_{Stroke}=92), the naturalness of the surface texture by the texture analysis module (TextureNet) is 88 points (S_{Texture}=88), and the aging match result by the crack analysis module (CrackNet) is 75 points (S_{Crack}=75), then the final image analysis score (S_{Image}) according to the above mathematical formula 1 is calculated as a total of 87.8 points through the calculation process of (0.6 * 92) + (0.2 * 88) + (0.2 * 75). Based on the 87.8 points quantified in this way, the processor identifies that the visual characteristics of the artwork correspond to a high probability of authenticity range, and integrates this with the results of the material and history analysis engine to determine the final integrated authenticity appraisal index.

[0083] The S_{Image} value generated in the above process is utilized as a key indicator representing the external legitimacy of the artwork, and contributes to securing the integrity of the final appraisal result through non-linear penalty logic and cross-validation procedures with other domain data performed by the processor.

[0084] This quantitative calculation method overcomes the limitations of conventional technology that relied on the subjective eye of appraisal committee members and provides the effect of dramatically improving the reliability of image-based visual evidence as a basis for art transactions and legal grounds by standardizing it into objective numerical values.

[0086] In the material analysis stage, X-ray fluorescence (XRF), Raman spectroscopy, and Fourier Transistor (FTIR) data are utilized to analyze the elemental composition, molecular structure, binder, and matrix characteristics of the pigments. In particular, the use of pigments that could not have existed during that period or artificial aging treatments are incorporated into the calculation process as decisive evidence of forgery. The Material AI receives spectral data values ​​obtained from precision analysis equipment through an input interface and performs matching with reference data. XRF data analysis identifies the elemental composition pattern of the pigments on the artwork's surface and compares it with the artist's unique pigment blending method, while Raman and FTIR analyses identify molecular and crystal structures to distinguish between natural and synthetic pigments and measure the degree of binder oxidation. If a critical age discrepancy signal is detected, such as when a specific pigment component developed after the mid-20th century is detected in a 19th-century work, the processor considers this a 'Fatal Flaw' and imposes a heavy penalty on the material analysis score (S_{Material}).

[0087] In the history analysis stage, document data extracted via OCR is analyzed to comprehensively quantify Document Integrity, Event Consistency, Historical Validity, and Forgery Trace. The Provenance AI engine converts document images—such as warrants, exhibition records, and auction records—into digital text and utilizes web crawling technology to verify consistency with external online databases. The processor analyzes the document's font, formatting, and paper noise patterns to detect traces of forgery, and derives a history reliability score by cross-validating the historical validity of the spatiotemporal path of the possession history claimed by the document.

[0088] The base weighted sum score (S_{base}), which serves as the basis for calculating the final score, is calculated by integrating the results of the three independent engines by the processor, and can be expressed as shown in the following mathematical formula 2.

[0089] [Mathematical Formula 2]

[0090]

[0091] W_I, W_M, and W_P represent dynamic weights assigned according to domain importance to scores derived from the image analysis engine (Image AI), material composition analysis engine (Material AI), and literature history analysis engine (Provenance AI), respectively. At this time, the sum of the weights is set to satisfy the condition 1 (W_I + W_M + W_P = 1) to normalize the numerical range of the basic weighted sum score. S_{Image}, S_{Material}, and S_{Provenance} represent independent authenticity scores calculated by each analysis engine.

[0092] Based on the calculated base weighted sum score (S_{base}), the processor determines the final integrated authenticity assessment index (IAI Score, hereinafter Score_{final}) by applying a non-linear penalty according to data conflict patterns found in a specific analysis area.

[0093] A specific method and numerical example for a processor according to one embodiment of the present invention to calculate each parameter included in Equation 2 and Equation 6 to calculate the final integrated authenticity assessment index, IAI Score, is as follows. The processor obtains independent authenticity scores (S_{Image}, S_{Material}, S_{Provenance}) derived from an image analysis engine, a material component analysis engine, and a literature history analysis engine, calculates a base weighted sum score (S_{base}) by applying weights (W_I, W_M, W_P) dynamically assigned according to the importance of each region to the scores, and then determines the final index by applying a non-linear penalty logic based on a forgery match probability (P_{match}) that reflects contradictions between data.

[0094] Specifically, the material component analysis score (S_{Material}) is calculated by directly comparing the peak values ​​and waveforms of the spectroscopic spectra output from XRF, Raman, and FTIR equipment with a standard database (Standard DB). The processor does not merely determine whether the components match, but also inputs historical storage environment information (temperature, humidity, cumulative UV exposure, etc.) restored from literature history into a simulation model to predict the chemical degradation of pigments and binders. Based on the aging consistency index (C_{age}) derived therefrom, it verifies the spatiotemporal causal relationship with actual measurement data to quantify the score. In particular, if a fatal flaw is identified, such as the detection of synthetic pigments that do not match the manufacturing date, the S_{Material} is rapidly reduced to the lowest range.

[0095] Furthermore, the Provenance score is derived by analyzing text data, such as digitized certificates, exhibition and auction records, using OCR technology. The processor calculates the score by synthesizing a Document Integrity score, which verifies the integrity of the document's font and formatting; an Event Consistency score, which confirms the existence of an event through cross-verification with an external database; and a Historical Validity score, which analyzes the consistency with the artist's life cycle and geographical movement path. The processor confirms the reliability of the score by verifying the compatibility between the submitted documents and data collected in real-time from online museum archives and public data portals via a web crawler.

[0096] The dynamic weights (W_I, W_M, W_P) used in the above mathematical formula 2 are adjusted in real time according to the type of artwork to be analyzed and the completeness of the input data. The sum of the weights always satisfies the normalization condition of 1, but if the data uncertainty (Aleatoric Uncertainty) parameter calculated by a specific engine exceeds a threshold or if a missing essential literature data is detected, the processor executes a dynamic weight redistribution logic that non-linearly attenuates the weight of the engine and increases the weight of a scientific analysis engine (Material AI, etc.) with relatively high reliability.

[0097] The forgery match probability (P_{match}) is a value imposed via Equation 4 when a logical contradiction or data conflict is detected between different analysis engines, and the processor generates a virtual forgery dataset (S_{fake}) containing the corresponding conflict pattern through a Generative Adversarial Network (GAN) or a diffusion model to analyze the similarity with the current object of appraisal. If the match probability (P_{match}) with the simulated forgery technique exceeds a threshold (T_{th}), the processor activates a multiplicative elimination logic based on a sigmoid function defined in Equation 6 to cause the final indicator (IAI Score) to drop sharply to the forgery judgment range.

[0098] As a specific numerical example regarding the calculation of the above IAI Score, we assume a case where a 19th-century oil painting with partially lost literature history is appraised. The processor recognizes the state of missing literature and adaptively redistributes weights as W_I=0.3, W_M=0.6, and W_P=0.1. At this time, it is assumed that in a situation where the individual scores derived from each engine are S_{Image}=85, S_{Material}=90, and S_{Provenance}=40, the material analysis results locally detect modern pigment components that could not have existed in the 19th century, and the forgery match probability (P_{match}) exceeds the penalty activation threshold. In this case, multiplicative elimination based on the penalty logic of mathematical formula 6 is applied, and regardless of the base sum score (25.5 + 54 + 4 = 83.5), the final indicator (Score_{final}), IAI Score, is reset by dropping sharply to the 'very high probability of forgery' range of less than 20 points. Through this non-linear scoring procedure, the present invention fundamentally blocks forgery attempts that aim to precisely imitate only specific domains and ensures the objective reliability of the appraisal results.

[0100] The aforementioned processor goes beyond simply calculating the average of the authenticity probabilities derived by each engine; it incorporates patterns of mutual reinforcement or conflict among evidence from three domains into the analysis model. For instance, even if image analysis indicates a high probability of authenticity, if pigments with an inconsistent dating are detected during material analysis, a penalty logic is activated to forcibly override the calculated integrated score to the forgery judgment range. This fundamentally blocks sophisticated forgery attempts designed to bypass specific analysis domains and ensures the integrity of the appraisal results.

[0101] The processor sets basic weights for each of the three domains of image, material, and history, and is configured to dynamically adjust these weights based on the type of artwork to be identified (e.g., oil painting, ink painting) or the presence or absence of literature. For example, in the case of an antique artwork lacking a literature history, the weight of the material analysis results is adaptively increased. In particular, if fatal evidence of forgery (Fatal Flaw), such as a discrepancy in production date, is found in a specific analysis area among the three domains, a non-linear scoring procedure is performed to cause the final index (IAI Score) to plummet to the forgery judgment range by applying a sharp subtraction function that acts independently of the weighted sum result, even if the analysis results of other domains support authenticity.

[0102] The processor calculates, in real time, an aleatoric uncertainty parameter—which indicates the confidence level of the result caused by noise or missing information in the input data—along with analysis results from each of the three independent artificial intelligence engines that perform image, material, and history analysis. For analysis results from engines where the aleatoric uncertainty exceeds a preset threshold, the processor is configured to non-linearly decay the weight of the result when calculating the final indicator. In particular, if a critical defect suggesting forgery is detected by a specific engine, a dynamic weight allocation procedure is performed to maintain the precision of the determination by redistributing weights independently of the analysis results of other engines.

[0103] The IAI Score, finally calculated by the processor, is systematically interpreted according to score ranges. The 90 to 100-point range signifies a state of very high authenticity probability, where the analysis results of all three AI engines strongly support authenticity; the 75 to 89-point range is defined as high authenticity probability, where most indicators align well with the characteristics of an authentic item. The neutral range of 50 to 74 points indicates a state where evidence of authenticity and elements suspected of forgery are mixed, while the 30 to 49-point range is interpreted as having low authenticity probability, as negative signals are detected by multiple AI modules. Finally, the 0 to 29-point range signifies a state of very high probability of forgery, where the main analysis patterns significantly inconsistent with authentic data. In this way, the processor operates to maximize the precision of the appraisal by integrating and resetting the scores of each independent engine according to a mutual verification logic.

[0105] In addition, an electronic device (110) and / or processor according to one embodiment of the present invention, through the following method, a quantitative authenticity assessment index (IAI Score) based on artificial intelligence (AI) TMIt is possible to calculate and / or reset ) or determine whether the artwork is authentic and generate authenticity determination information, which is information regarding integrated authenticity determination of artwork based on artificial intelligence, and to guide the user regarding the authenticity of the artwork through the user's terminal (120).

[0106] The processor is configured to perform re-analysis by reflecting the analysis data of other engines as conditional parameters in the relevant conflict area when a conflict pattern is detected in which the results do not match between the analysis engines, and to non-linearly downward reset the integrated authenticity assessment index by applying a preset penalty logic if circumstances of forgery are confirmed as a result of the re-analysis.

[0107] Specifically, the processor calculates an aleatoric uncertainty parameter, which indicates the confidence level of the result, along with the analysis result value from each of the three independent artificial intelligence engines that perform image, material, and history analysis in real time, and automatically reduces the weight reflected in the calculation of the integrated authenticity assessment index for the analysis result of an engine in which the aleatoric uncertainty exceeds a preset threshold, while performing a dynamic weight allocation procedure that independently controls the weight of other engines when a critical defect is detected in a specific engine.

[0108] If a conflict pattern is detected in which a contradiction in result values ​​occurs between multiple analysis engines, the processor is configured to extract the feature point where the conflict occurred on the image data as a local ROI, and to perform a conditional attention-based reanalysis on the local ROI in which material component numerical values ​​or literature history data, which are analysis results of other engines, are reflected as conditional parameters, thereby finally determining whether to apply the penalty logic.

[0109] Furthermore, the processor is characterized by maximizing the precision of authentication through a cross-verification process that extracts information on the past storage environment of the artwork from the aforementioned literature history data to simulate the chemical alteration of pigments and binders, and dynamically corrects the reference data of the material analysis engine to determine whether the detected conflict is due to natural aging or artificial alteration. The operation of this processor fundamentally prevents the possibility of bypassing conventional technology that relies on a single analysis method, and ensures a high degree of advancement by logically combining causal relationships between data from different domains to derive quantitative indicators.

[0111] The processor is configured to obtain analysis results (e.g., probability of authenticity) for artworks from each of the three independent artificial intelligence engines, namely the image analysis engine, the material composition analysis engine, and the literature history analysis engine, and to calculate in real time an aleatoric uncertainty parameter that quantitatively represents the confidence level of the result resulting from noise in the input data itself, measurement error, or incompleteness of information. To this end, the processor derives the aleatoric uncertainty parameter by performing a computational procedure to calculate the variance of the output distribution by applying a Bayesian approach during the inference process of each artificial intelligence engine or by performing multiple Monte Carlo Dropout samplings.

[0112] The processor executes a reliability-based weight adjustment logic comprising the steps of identifying a specific artificial intelligence engine whose calculated data uncertainty parameter exceeds a preset reliability threshold, and non-linearly decaying the base weight applied to the analysis result of the specific artificial intelligence engine in a multivariate weighted sum model for calculating the final integrated authenticity sentiment index (IAI Score), based on the magnitude of the data uncertainty parameter. In other words, it is a structure that automatically reduces the influence of the data on the final judgment as the reliability of the data decreases.

[0113] Furthermore, the processor performs an adaptive routine that dynamically controls weights by considering mutually complementary data characteristics. Specifically, if a state of missing data is detected in the historical analysis results—such as the omission of essential holding records, an exceeding ratio of areas where OCR recognition is impossible, or a failure of cross-verification with an external database—the processor recognizes that the relative importance of the material composition analysis results, which are objective scientific data, has increased. Accordingly, the processor performs a dynamic weight redistribution process that actively adjusts the weight assigned to the material composition analysis engine upward according to a preset upward ratio or upward function, independently of whether other engines' weights fluctuate. This is an advanced integrated control method that actively compensates for the absence of data in a specific area with highly reliable data from other areas, and it possesses significant advancement compared to a simple average calculation method.

[0115] The above processor can execute a time-series aging simulation model that predicts physical changes over time from the time of production to the present, using the oxidation degree of the binder and discoloration data of the pigment derived from the material analysis data as input.

[0116] The processor generates virtual crack pattern data by numerically simulating the crack occurrence and propagation process according to the passage of time and environmental variables (temperature, humidity, UV exposure) through the time-series aging simulation model. Subsequently, the processor performs a structural similarity calculation procedure that precisely compares the virtual crack pattern data with actual detected crack pattern data, which is extracted from high-resolution surface images of the actual artwork by the crack analysis module (CrackNet) and modeled in graph form, at the pixel level or graph node level. At this time, the processor is configured to calculate the structural similarity multidimensionally by analyzing the correlation between micromorphological parameters, such as crack bifurcation density, segment length distribution, and crack boundary roughness, going beyond simple visual shape comparison.

[0117] The processor identifies that if the calculated structural similarity is below a preset threshold for genuine aging, the actually detected crack pattern is an atypical pattern that is difficult to form over the natural passage of time. Accordingly, the processor determines that the currently detected aging state corresponds to a scenario of 'counterfeiting by artificial aging,' such as chemical treatment or the application of physical stress, and performs a series of control operations to immediately activate the penalty logic that applies a strong subtraction to the integrated authenticity assessment index, even if the analysis results of other artificial intelligence engines support genuineness. This operation of the processor is an advanced method that cross-verifies static material analysis data and dynamic simulation prediction data, and is an advanced technology that scientifically identifies sophisticated counterfeiting methods that cannot be detected by simple component analysis alone.

[0119] The processor can extract a feature point where the collision occurred on the image data when any one of the three independent artificial intelligence engines detects a collision pattern suggesting the possibility of forgery.

[0120] The processor can control the artificial intelligence engine performing the image analysis to perform a conditional attention-based reanalysis reflecting the analysis results (material or history) of another engine on a local region (Local ROI) containing the extracted feature points, thereby finally determining whether to apply the penalty logic.

[0121] The processor can dynamically correct the reference data of the material analysis engine by extracting information on the past storage location and environment of the artwork (at least one of temperature, humidity, and illuminance) from the literature history data and performing a simulation of chemical alteration of pigments and binders based on the extracted environmental information. Specifically, the processor identifies the geographical location where the artwork stayed and the duration of storage at that location from provenance and exhibition record texts digitized via OCR using Named Entity Recognition technology, and performs a preprocessing procedure to obtain data on the annual average temperature, relative humidity distribution, and cumulative UV exposure for the corresponding coordinates and specific periods by linking this with an external Historical Climate DB through a communication module. Subsequently, the processor inputs the acquired environmental variables as parameters of a physico-chemical aging model to numerically predict the degree of molecular structural alteration and the resulting changes in spectroscopic properties caused by oxidation, hydrolysis, and photolysis reactions that specific pigment components and organic binders identified as used in the artwork undergo over decades to centuries. Through this, the material analysis engine generates corrected reference data in which wavelength-specific intensity and peak shifts are calculated by reflecting the unique 'environmental history' that only the artwork has undergone, instead of static reference spectrum data applied uniformly to all artworks, and utilizes this as a control group against actual measurement data.

[0122] The processor can determine whether the conflict with the image analysis result is due to 'natural aging' or 'artificial tampering' based on the deviation between the corrected reference data and the actual measured material analysis data. In particular, the processor does not stop at merely determining whether the currently measured component values ​​match, but calculates an 'aging consistency index (C_{age})' defined as in Equation 3 below, and performs a calculation step to quantitatively verify the spatiotemporal causal relationship between the storage history described in literature and the physical and chemical aging state possessed by the actual artwork.

[0123] [Mathematical Formula 3]

[0124]

[0125] To explain in detail each parameter used in the above mathematical formula 3, S_{raw}(λ) is wavelength-specific spectral intensity data actually measured at a specific point on the artwork using XRF, Raman, or FTIR equipment, and S_{ref}(λ) is standard reference spectral data before denaturation of pigments and binders presumed to have been standardly used at the time of production of the artwork. Φ(T, H, L) is a 'cumulative environmental denaturation function' calculated by taking the average temperature (T), relative humidity (H), and cumulative UV exposure (L), extracted from literature history by the processor, as independent variables. ω(λ) is a wavelength-specific weight set to increase the reliability of the analysis for a specific wavelength band where the aging characteristics of a specific pigment appear most sensitively, and n is the number of sampling points in the entire wavelength band subject to comparative analysis.

[0126] The above processor calculates the aging consistency index (C_{age}) described in Equation 3 based on past storage environment information extracted from the above literature history data, and the specific calculation and operation process for each parameter is as follows.

[0127] S_{raw}(λ) included in the dividend term of the above mathematical formula 3 is wavelength (λ) spectral intensity data measured in real time at a specific point of the artwork through XRF, Raman, or FTIR equipment, and the processor is configured to acquire the digital raw data from an external analysis device through a communication module. S_{ref}(λ) is the standard reference spectral data before denaturation of pigments and binders that were predefined as being standardly used at the time of production of the artwork, and this is stored and managed by wavelength band in the reference database (Reference DB) within the memory. The cumulative environmental denaturation function Φ(T, H, L) is a physicochemical denaturation coefficient calculated by taking the average temperature (T), relative humidity (H), and cumulative ultraviolet exposure (L) according to the past storage location and period of the artwork, extracted from the literature history by the processor through OCR and named entity recognition technology, as independent variables. The processor numerically predicts the degree of molecular structure alteration over time from the time of production to the present by inputting time-series environmental variables obtained in conjunction with the external meteorological history database into oxidation, hydrolysis, and photolysis reaction models. The wavelength-specific weighting factor ω(λ) is a pre-set weighting factor to strategically increase the reliability of the analysis for a specific wavelength band (e.g., a specific infrared wavelength band with a distinct peak shift due to oxidation reaction) where the aging characteristics of a specific pigment or binder appear most sensitive to chemical variation, and n represents the number of sampling points in the entire wavelength band set to ensure statistical significance.

[0128] An example of the specific computational process and actual numerical input of the processor for deriving the aging consistency index (C_{age}) is as follows. It is assumed that the artwork to be appraised is an oil painting produced in the mid-19th century (about 170 years ago), and that a literature history analysis confirms that the work has been stored for a long period in a specific art museum in Europe. Based on the weather records of the region, the processor derives environmental parameters such as the average temperature T = 15 °C, relative humidity H = 65 %, and cumulative UV exposure L = 1,200,000 kJ / m2 during the storage period. Subsequently, the processor drives the cumulative environmental degradation function Φ(15, 65, 1.2M) for the organic binder component to predict that the specific peak intensity of the standard reference spectrum (S_{ref}(λ)) will attenuate by about 15% over time and the wavelength will shift slightly.

[0129] For example, if the actual data at a specific wavelength (λ_1) is S_{raw}(λ_1) = 0.82, the uncorrected standard reference data is S_{ref}(λ_1) = 1.0, and the aging sensitivity weight is ω(λ_1) = 1.5, the error term at that point is calculated as |0.82 - (1.0 * 0.85)| / 1.0 = 0.03. The processor calculates C_{age} by summing and averaging the weighted deviations above for a total of n sampling points.

[0130] If the calculated C_{age} value is within a preset confidence limit (e.g., 0.05 or less), the processor determines that the chemical aging state of the artwork currently possessed is a state of 'natural aging' that logically perfectly matches the archival records, and maintains the authenticity score. Conversely, if the actual measured oxidation degree of the pigment is significantly higher or lower than the simulation prediction value, causing C_{age} to exceed a threshold, the processor identifies this as an 'artificial tampering' scenario caused by forced aging using chemicals or forgery of the archival history. In this way, the processor generates dynamic correction spectrum data that reflects the unique environmental history of the artwork itself, instead of static reference data, and compares it with the actual measurement data; thereby scientifically blocking sophisticated forgery attempts that merely mimic component values ​​through the analysis of chemical causality along the time axis, and dramatically improving the precision of the authentication.

[0131] The processor determines that if the aging consistency index (C_{age}) calculated by the above mathematical formula 3 is within a preset confidence interval, the microcracks on the surface or discoloration of the pigment detected by the image analysis engine (e.g., CrackNet) are the result of 'natural aging' that may occur under the storage location and environment recorded in the literature. On the other hand, if the index exceeds a threshold, resulting in a statistically significant deviation between the actual data and the simulation results, and at the same time, if the image analysis engine detects unnaturalness in the brushwork or crack patterns caused by artificial heating or chemical treatment, the processor determines this as an 'artificial tampering' scenario attempted to mimic only specific analysis indicators. The operation of this processor verifies the logical consistency between external dynamic data, such as 'literature-based historical storage records,' and the physical entity, such as 'current material analysis results,' through simulation. This overcomes the limitations of conventional technology that relies simply on static pigment DB comparison, and secures a high level of precision in identifying forgeries by utilizing the technical principle that even if only specific pigment components are precisely imitated, the causal relationship of aging along the time axis cannot be forged. Unlike prior art that merely compares static material values, this has the technical distinctiveness of dramatically improving the precision of authentication through dynamic data correction along the time axis.

[0132] When a collision occurs between the analysis data, the processor may generate virtual forgery data having the corresponding collision pattern through a Generative Adversarial Network (GAN) or a Diffusion Model using the acquired data as input. When mutual collision between the acquired analysis data is detected, the processor performs the step of generating multiple virtual forgery data having the corresponding collision pattern by driving a Generative Adversarial Network (GAN) or a Diffusion Model that receives a specific feature where the collision occurred as a conditional parameter. Here, the term "mutual collision" refers to a state in which visual features derived from an image analysis engine support the pattern of an authentic product from a specific era, but physical component data derived from a material analysis engine includes a material that does not exist in the specific era, or lacks a spatiotemporal causal relationship with the results of the history analysis.

[0133] Specifically, the processor monitors the logical consistency between feature vectors output in real time from each independent analysis engine (Image AI, Material AI, Provenance AI) and defines specific feature combinations where the deviation between data exceeds a threshold based on a pre-configured correlation map as 'potential forgery signals'. The processor feeds the identified potential forgery signals into the conditional layer of the generative model to simulate 'combinations of contradictory feature points' that do not exist in the actual authentic dataset but may occur during the production of forgeries. For example, the processor sets a condition (C) as an input value where the pressure pattern of the brushstrokes (image feature) mimics the habits of an 18th-century artist, but the chemical concentration of the pigment used (material feature) matches 20th-century synthesis methods. Subsequently, the processor performs a procedure to dynamically construct a set of tens of thousands of virtual forgery samples (S_{fake}) with logical defects similar to the current object of appraisal by restoring the data distribution that can be generated under these contradictory conditions through adversarial learning or a reverse diffusion process.

[0134] Unlike prior art, which remained a passive method of merely calculating the match rate with a static database of genuine works (DB), the operation of this processor achieves a high degree of advancement in that it visualizes the technical reality of forgery by inversely tracing, through generative artificial intelligence, 'inconsistencies between data' that a forger may intentionally manipulate or accidentally leave behind. In particular, the processor is configured to determine whether feature data (X_{target}) extracted from the artwork currently under appraisal exists within the statistical distribution area of ​​a virtual forgery dataset (S_{fake}) simulated by a generative model through the calculation of a 'forgery match probability' according to Equation 4 below.

[0135] [Mathematical Formula 4]

[0136]

[0137] The technical significance of each parameter included in the above mathematical formula is explained in detail as follows. First, P_{match} (forgery match probability) is a probability value between 0 and 1 representing the degree to which the artwork under appraisal matches a hypothetical forgery scenario simulated by the generative model. As the P_{match} value converges to 1, the artwork is judged to perfectly match the logical flaws of a specific forgery technique predicted by the artificial intelligence; to this end, the correction constant k_{scale} is normalized to have a range greater than 0 and less than or equal to 1 (0 < k_{scale} ≤ 1). Next, D_{dist} (data distance) refers to the Euclidean distance or cosine similarity-based distance between the centroid of the feature distribution of the hypothetical forgery dataset formed in the generative model's high-dimensional latent space and the feature vector (V_{target}) of the artwork currently under appraisal. As the artwork under appraisal is located within the statistically dense region of the forgery technique, the D_{dist} value converges to 0. In addition, k_scale (correction constant) is a weighting factor that is adaptively set according to the number of dimensions of the data and the reliability parameter of the analysis engine where the conflict occurred. It serves to optimize the sensitivity of forgery detection according to the severity of the conflict when a contradiction between specific domains is detected, and the 1 in the denominator (stabilization term) is a fixed constant introduced to prevent computational errors caused by the denominator becoming zero and to normalize the final result value to 1 or less.

[0138] The processor performs a penalty control operation in which, if the forgery match probability (P_match) calculated by the above mathematical formula 4 exceeds a preset penalty activation threshold, it immediately invalidates the image analysis results, etc., even if they correspond to the genuine range, and forcibly resets the Integrated Authenticity Assessment Index (IAI Score™) to the forgery grade. This procedure of verifying forgery scenarios by backtracking minute collision patterns occurring between data of different characteristics using a generative model provides a technical effect that dramatically improves the precision of assessment by fundamentally blocking forgery attempts that attempt to sophisticatedly bypass analysis of a single area. Consequently, the present invention overcomes the limitations of existing single analysis methods and maximizes the reliability of assessment results by systematically designing logical causal relationships between data to derive quantitative indicators.

[0139] A processor according to one embodiment of the present invention performs the step of generating a plurality of virtual forgery data having a corresponding conflict pattern by driving a generative model that receives a specific feature where the conflict occurred as a conditional parameter when a mutual conflict is detected between the acquired analysis data. The processor is configured to precisely compare the generated virtual forgery data with the artwork data currently under appraisal to calculate the probability that the detected peculiar pattern matches the statistical characteristics of a previously known forgery technique.

[0140] Specifically, the processor quantifies how closely the object of appraisal matches the statistical characteristics of the simulated forgery technique by calculating the distance between the multidimensional feature vector (V_{target}) extracted from the artwork of appraisal and the feature distribution center point (V_{fake}) of the virtual forgery dataset generated through the simulation in a high-dimensional latent space as shown in Equation 5 below.

[0141] [Mathematical Formula 5]

[0142]

[0143] (where D_dist represents the numerical distance between the feature vector of the artwork to be appraised (V_{target}) and the feature vector of the virtual forgery (V_{fake}), where V_{target} is the feature vector of the artwork to be appraised and V_{fake} is the feature vector of the virtualally generated forgery.)

[0144] In the above mathematical formula 5, V_{target} is defined as a 'Fused Feature Vector' formed by normalizing and combining the stroke feature of the image analysis engine, the material feature of the material analysis engine, and the provenance integrity of the history analysis engine.

[0145] The processor quantifies the numerical distance between the actual data of the artwork to be appraised and the simulated forgery dataset in a high-dimensional latent space through the above mathematical formula 5, and the specific calculation method and definition of the mathematical symbols for each parameter included therein are as follows. V_{target} in the above mathematical formula 5 is a multidimensional feature vector extracted from the artwork to be appraised, and is defined as a 'Fused Feature Vector' formed by normalizing the Stroke Feature obtained through the image analysis engine, the Material Feature obtained through the material analysis engine, and the Provenance Integrity obtained through the history analysis engine, respectively, and then combining them into one. The processor generates the above V_{target} by projecting the data of different dimensions derived by each engine onto a common latent space. The above V_{fake} represents the centroid of the feature distribution of a virtual forgery dataset simulated by a Generative Adversarial Network (GAN) or a Diffusion Model, and is calculated as a vector value that statistically represents the combination of contradictory feature points that may occur during the forgery production process.

[0146] The double bracket symbol (||·||) used in the above mathematical formula 5 represents a 'norm,' an operator that measures the magnitude or length of a vector, and the number '2' written at the bottom right indicates that the 'L2 norm' is specifically applied. This refers to the Euclidean distance measurement method, which calculates the shortest straight-line distance between two points in a multidimensional space, and is calculated by taking the square root of the sum of the squares of the differences between each component. D_{dist}, derived through this calculation procedure, is used as a high-dimensional similarity indicator to show how close the artwork being appraised is to the statistically dense area of ​​a specific forgery scenario predicted by artificial intelligence.

[0147] As a specific numerical example for calculating the above D_{dist}, the computational process for a forgery scenario suspected of imitating an 18th-century antique art piece is as follows. It is assumed that the value of the fused feature vector V_{target} extracted from the artwork subject to appraisal by the processor is defined as [0.85, 0.92, 0.10] in a normalized 3D space, and that the center point V_{fake} of the forgery dataset simulated by the generative model is calculated as [0.84, 0.20, 0.12]. At this time, the processor calculates the deviations for each element [0.01, 0.72, -0.02], calculates the sum of the squares of each deviation, 0.0001 + 0.5184 + 0.0004 = 0.5189, according to the L2 norm formula, and then takes the square root of this to finally derive D_{dist} - 0.7203.

[0148] The data distance (D_{dist}) calculated above is subsequently used as an input parameter in the calculation process of Equation 4 to determine the final forgery match probability (P_{match}). If, as in the above embodiment, a significant deviation occurs in a specific domain (e.g., material composition) and the D_{dist} value converges close to 0, the processor determines that the artwork perfectly matches the statistical characteristics of the forgery technique predicted by the generative model. This calculation procedure scientifically blocks forgery attempts that attempt to sophisticatedly bypass analysis of a single domain through distance analysis in a high-dimensional latent space, and provides the effect of maximizing the objective reliability of the appraisal result by quantifying logical defects between data.

[0149] The above processor calculates the forgery technique match probability (P_{match}) based on the calculated distance between features (D_dist), which is used as an input parameter in the operation process of Equation 6 below.

[0150] The above processor is configured to calculate the final integrated authenticity assessment index (Score_{final}) according to the following mathematical formula 6 by applying a 'Multiplicative Score Nullification Logic' based on a sigmoid function based on the forgery match probability (P_{match}) to the weighted sum result of individual scores (S_i) derived from three independent artificial intelligence engines.

[0151] [Mathematical Formula 6]

[0152]

[0153] Score_{final} represents the final confirmed integrated authenticity assessment index.

[0154] k_{steep} represents the penalty steepness parameter that determines the severity of the reduction when forgery circumstances are detected.

[0155] Here, P_{match} represents the forgery match probability derived from Equation 4, T_{th} represents the activation threshold at which the penalty logic is activated, and k_{steep} represents the penalty steepness parameter that determines the severity of the score reduction when forgery circumstances are detected.

[0156] Through this non-linear structure, the final appraisal index drops sharply exponentially the moment the probability of forgery exceeds a specific threshold, thereby maximizing the ability to distinguish between genuine and counterfeit works.

[0157] The processor applies a penalty term in the form of a sigmoid function, but controls the value of the penalty term to rapidly converge from 0 to 1 as P_{match} exceeds an activation threshold (T). Accordingly, the attenuation coefficient multiplied by the base sum score converges to 0, thereby implementing non-linear behavior in which the final indicator (Score_{final}) rapidly collapses into the forgery judgment range regardless of whether the item supports authenticity in a specific area. Here, the parameter k_steep determines the severity of the deduction applied to the appraisal score when forgery circumstances are detected, and the processor can readjust the penalty strength by adaptively adjusting the k_steep value upward or downward according to the historical significance of the appraisal subject or data uncertainty parameters.

[0158] Furthermore, the processor inputs historical storage environment time-series information (average temperature T, relative humidity H, cumulative UV exposure L) restored from literature history data as parameters for a physicochemical aging model to numerically simulate the aging of pigments and organic binders. Based on the aging consistency index (C_{age}) calculated through this process, static reference spectrum data is converted into a 'dynamically corrected spectrum' that reflects the unique history of the artwork, and this is used as a control group against the measured data.

[0159] [Mathematical Formula 7]

[0160]

[0161] (where S_{raw}(λ) is the measured spectral intensity, S_{ref}(λ) is the standard reference spectrum, Φ(T, H, L) is the cumulative environmental degeneration function, and ω(λ) is the wavelength-specific weight.)

[0162] The processor determines whether the surface crack pattern detected by the image analysis engine (CrackNet) is naturally occurring under the storage environment based on the aging consistency index (C_{age}) calculated through Equation 7. If the calculated aging consistency index (C_{age}) deviates from a preset confidence limit and a statistically significant deviation occurs between the simulation prediction and the actual data, the processor confirms this as an 'artificial tampering' scenario attempted to disguise it as natural aging. At this time, the processor executes a sequential control routine to nullify the final indicator by activating the non-linear penalty logic of Equation 6 by forcibly updating the forgery match probability (P_{match}) of Equation 4 above a threshold (T).

[0163] As such, unlike conventional technology that was limited to static data comparison, the present invention dramatically improves the precision of the emotion by simulating contradictions between data of different domains using a generative model and verifying them by combining them with chemical aging logic along the time axis.

[0164] The embodiments disclosed in this specification and drawings are merely examples to aid in understanding the technology and do not limit the scope of the invention; various modifications and combinations of embodiments based on the technical concept are possible.

[0165] The device and method according to the present invention may be realized in hardware, software, or a combination thereof, and may be recorded in the form of program instructions on various computer-readable media such as ROM, RAM, optical / magnetic disks.

[0166] Such programs may be distributed and stored over a network or electronically transmitted via communication signals, and the present invention includes both code implementing the device or method described in the claims and a medium storing the same. Accordingly, the true scope of protection of the present invention must be determined by the claims.

[0167] delete

Claims

Claim 1 An electronic device for determining the authenticity of a work of art, comprising: a memory for storing data regarding a work of art; The system includes a processor that executes instructions stored in the memory, wherein the processor acquires image data, material composition analysis data, and literature history data of the artwork, individually analyzes the acquired data through three independent artificial intelligence engines that respectively perform image analysis, material analysis, and history analysis, and reflects patterns in which the respective analysis evidence derived from the three independent artificial intelligence engines reinforces or conflicts with one another in an analysis model to calculate an integrated authenticity appraisal index that quantifies the probability of the artwork's authenticity as a numerical value between 0 and 100, wherein the processor performs preprocessing on the image data including noise removal, color normalization, and illumination fix, and divides the preprocessed image data into patches of 128, 256, or 512 pixels (px) to set zones for brushstroke and crack analysis, wherein the artificial intelligence engine performing the image analysis performs brushstroke analysis that verifies whether the artist's unique brushstroke grammar matches based on brushstroke flow, curvature, and pressure maps. Module (StrokeNet); a texture analysis module (TextureNet) that detects periodicity and unnatural uniformity of surface textures using deep features based on LBP (Local Binary Patterns) and CNN (Convolutional Neural Network); and a crack analysis module (CrackNet) that models the cracking structure of paint layers and determines whether cracks are artificial using U-Net and GNN (Graph Neural Network);The AI ​​engine performing the material analysis is configured to receive data values ​​output from at least one of X-ray fluorescence analysis (XRF), Raman spectroscopy (Raman), and Fourier Transistor (FTIR) equipment, and to detect indirect evidence of forgery by comparing and analyzing the elemental composition, molecular structure, characteristics, and degree of aging of the pigment with reference data; the AI ​​engine performing the history analysis digitizes the document history data through OCR (Optical Character Recognition) technology and derives indicators of Document Integrity (comparing the font and format of the document), Event Consistency (analyzing cross-verification with an external database), and Historical Validity (analyzing the suitability with the artist's date); the processor resets the integrated authenticity assessment indicator by applying a pre-set penalty logic to the conflicting patterns if pigment components inconsistent with the production date are detected in the material analysis results or traces of document tampering are detected in the history analysis results, even if the image analysis results support authenticity; and the three independent AI engines An electronic device characterized by being configured to calculate an aleatoric uncertainty parameter indicating the confidence level of the result along with the analysis result value from each, and to automatically reduce the weight reflected in the calculation of the integrated authenticity appraisal index for analysis results of an engine where the aleatoric uncertainty exceeds a preset threshold, while independently adjusting the weight of the material analysis engine upward when a missing document is detected in the history analysis result. Claim 2 delete Claim 3 delete Claim 4 delete Claim 5 An electronic device according to claim 1, wherein the processor executes a time-series aging simulation model that predicts physical changes over time from the time of production to the present, using the oxidation degree of the binder and discoloration data of the pigment derived from the material analysis data as input; if the structural similarity between the virtual crack pattern predicted by the simulation model and the crack pattern actually detected by the crack analysis module (CrackNet) is less than a threshold value, it determines this as a 'counterfeiting by artificial aging' scenario and activates the penalty logic; and if any one of the three independent artificial intelligence engines detects a collision pattern suggesting the possibility of counterfeiting, it extracts a feature point where the collision occurred on the image data; and controls the artificial intelligence engine performing the image analysis to perform a conditional attention-based re-analysis reflecting the analysis results (material or history) of other engines on a local ROI containing the extracted feature point, thereby finally determining whether to apply the penalty logic. Claim 6 In paragraph 5, the processor extracts past storage location and environmental information of the artwork from the literature history data, wherein the environmental information includes at least one of temperature, humidity, and illuminance; performs a simulation of chemical alteration of pigments and binders according to the extracted environmental information to dynamically correct the reference data of the material analysis engine; determines whether the conflict with the image analysis result is due to 'natural aging' or 'artificial alteration' based on the deviation between the corrected reference data and the actual measured material analysis data; if a conflict occurs between the analysis data, generates virtual forgery data having the corresponding conflict pattern through a Generative Adversarial Network (GAN) or a Diffusion Model using the acquired data as input values; compares the virtual forgery data with the artwork data currently under appraisal to calculate the probability that the conflict pattern matches the statistical characteristics of previously known forgery techniques; and sets the penalty logic to the maximum only when the calculated probability exceeds a preset threshold. An electronic device characterized by being configured to be applied.

Images