Art asset value evaluation method and system using artificial intelligence

Abstract

This art asset value evaluation system using artificial intelligence may comprise: a memory for storing instructions; and a processor. When executed by the processor, the instructions can instruct the system to: collect, from a plurality of data sources, artwork-related data including artist information, artwork characteristic information and market transaction information; normalize the collected data to generate learning data; use the learning data to train the current value prediction model, a price change prediction model, and an investment value evaluation model; use the trained models to calculate the current market value, future price change prediction, and a long-term investment value of a specific artwork; and integrate the calculated results to generate a final evaluation result.

Description

Method and System for Valuing Art Assets Using Artificial Intelligence

[0001] The present invention relates to a system and method for evaluating the value of art assets using artificial intelligence technology. Specifically, the invention relates to a system that analyzes information related to artworks collected from various data sources using artificial intelligence technology to predict the current market value and future price fluctuations of the artworks and to evaluate their long-term investment value.

[0002] The global art market has recently been experiencing rapid growth driven by the emergence of new forms of art, such as NFTs (Non-Fungible Tokens). While investment in art has become more active alongside this market expansion, both artists and investors are facing difficulties due to the lack of a systematic valuation system. For example, the case of digital artist B-ple's NFT artwork, which experienced a sharp decline in value after being auctioned for approximately 69.3 billion won in 2021, clearly illustrates the investment risks that can result from the absence of an objective valuation system.

[0003] In the current art market, artists at all levels—from emerging to mid-career and premium—face difficulties in receiving accurate valuations for their works. This not only hinders artists' sustainable creative activities but also poses unnecessary risks for investors. In particular, there is a problem in that existing art valuations rely heavily on the subjective judgment of experts, meaning the value of a work can vary significantly depending on the appraiser.

[0004] The introduction of a systematic valuation system holds significant meaning from an ESG perspective. This is because it can guarantee the economic sustainability of artists (Social), enhance the transparency and reliability of the art market (Governance), and promote the healthy development of the culture and arts ecosystem (Environmental). Furthermore, an objective valuation system can contribute to increasing the liquidity of artworks to stimulate trading and minimize the risk of loss for investors.

[0005] The present invention aims to solve the aforementioned problems by providing a system capable of objectively and quantitatively evaluating the value of artworks using artificial intelligence technology.

[0006] The present invention aims to accurately calculate the current market value by comprehensively analyzing various data, such as the physical characteristics of the work, the artist's history, and market trends.

[0007] Furthermore, the present invention aims to support investment decision-making by predicting future value fluctuations of a work through time series analysis, and to evaluate long-term investment value by considering the scarcity of the work and the growth potential of the artist.

[0008] An art asset valuation system using artificial intelligence may include a memory and a processor for storing instructions. The instructions, when executed by the processor, can control the system to collect art-related data including artist information, artwork characteristic information, and market transaction information from multiple data sources, normalize the collected data to generate training data, train a current value prediction model, a price fluctuation prediction model, and an investment valuation model using the training data, calculate the current market value, future price fluctuation prediction, and long-term investment value of a specific artwork using the trained models, and integrate the calculated results to generate a final evaluation result.

[0009] The art asset valuation system described in this document can enhance the transparency and reliability of the art market through data-driven, objective valuation. Furthermore, it can reduce investor risk and support rational decision-making by predicting future price fluctuations.

[0010] The art asset valuation system according to this document can promote diversity in the art market by providing fair evaluation opportunities for emerging artists and works in non-mainstream genres, and can present an objective evaluation system utilizing data and artificial intelligence technology, moving away from existing expert-centered evaluation methods.

[0011] FIG. 1 is a diagram illustrating an art asset valuation system using artificial intelligence according to one embodiment.

[0012] FIG. 2 is a diagram illustrating the learning of a neural network according to one embodiment.

[0013] FIG. 3 is a diagram illustrating the configuration of an artificial intelligence model according to one embodiment.

[0014] Figure 4 is a block diagram showing the configuration of an art asset valuation system using artificial intelligence according to one embodiment.

[0015] FIG. 5 is a flowchart illustrating a method for evaluating the value of art assets using artificial intelligence according to one embodiment.

[0016] FIG. 6 is a flowchart illustrating a method for evaluating the value of art assets using artificial intelligence according to one embodiment.

[0017] FIG. 7 is a flowchart illustrating a method for evaluating the value of art assets using artificial intelligence according to one embodiment.

[0018] FIG. 8 is a flowchart illustrating a method for evaluating the value of art assets using artificial intelligence according to one embodiment.

[0019] FIG. 1 is a diagram illustrating an art asset valuation system using artificial intelligence according to one embodiment.

[0020] As illustrated in FIG. 1, an art asset valuation system (100) using artificial intelligence may include a plurality of user terminals (110-1,…), a server (120), and a database (130). According to one embodiment, the database (130) is shown as being configured separately from the server (120), but is not limited thereto, and the database (130) may be provided within the server (120). For example, the server (120) may include a plurality of artificial intelligences for performing machine learning algorithms. According to one embodiment, the plurality of user terminals (110-1,…), the server (120), and the database (130) may be connected to communicate with each other through a network (N).

[0021] The network (N) can perform communication between multiple user terminals (110-1,…), a server (120), a database (130), etc. The network can perform wireless communication such as LTE, CDMA, WiFi, Bluetooth, NFC, GPS, etc. or wired communication such as USB, HDMI, RS-232, and the database (130) can store data acquired, processed, and used by these components.

[0022] In this invention, Artificial Intelligence (AI) refers to a technology that implements human learning ability, reasoning ability, and perceptual ability using a computer, and Machine Learning refers to an algorithmic technology that performs judgment or prediction by independently analyzing and learning the characteristics of input data. An AI learning model can be designed by mimicking the structure of the human brain, and multiple network nodes can exchange data based on convolutional connection relationships while located in layers of different depths. Such a model can be trained using methods such as supervised learning, unsupervised learning, and reinforcement learning, and algorithms such as decision trees, Bayesian networks, support vector machines, and artificial neural networks can be used. In particular, a Convolutional Neural Network (CNN) is a type of multi-layer perceptron that uses minimal preprocessing; it consists of one or more convolutional layers and artificial neural network layers, and utilizes weights and integration layers. CNNs can effectively process input data with a two-dimensional structure and demonstrate excellent performance in the fields of image and voice.

[0023] FIG. 2 is a diagram illustrating the learning of a neural network according to one embodiment.

[0024] As illustrated in FIG. 2, the learning device can train a neural network (123) to process review responses received from a plurality of user terminals (110-1,…) by item. Additionally, the learning device can train a neural network (123) to extract user stay history from user movement path information. According to one embodiment, the learning device may be a separate entity from the server (120), but is not limited thereto.

[0025] The neural network (123) includes an input layer (121) into which training samples are input and an output layer (125) that outputs training outputs, and can be trained based on the difference between the training outputs and the labels. Here, the labels are defined based on items corresponding to review responses and can be defined based on user dwell history corresponding to movement path information. The neural network (123) is connected as a group of multiple nodes and is defined by weights between the connected nodes and an activation function that activates the nodes.

[0026] FIG. 3 is a diagram illustrating the configuration of an artificial intelligence model according to one embodiment.

[0027] An artificial intelligence model according to one embodiment may include an input layer, a hidden layer, and an output layer.

[0028] The input layer is the layer associated with the input values ​​fed into the artificial intelligence model. In the hidden layer, feature maps can be output by performing MAC (multiply-accumulate) operations and activation operations on the input values. A MAC operation can be an operation that multiplies each input value by its corresponding weight and sums the multiplied values.

[0029] An activation operation may be an operation that inputs the result of a MAC operation into an activation function to output a result value. The activation function may be of various types. For example, the activation function may include a sigmoid function, a tangent function, a ReLU function, a Leaky ReLU function, a MaxOut function, and / or an ELU function, but there are no restrictions on the types. An output layer may be a layer associated with the result value of an operation performed in a hidden layer.

[0030] Figure 4 is a block diagram showing the configuration of an art asset valuation system using artificial intelligence according to one embodiment.

[0031] A system (400) according to one embodiment may include a processor (420) and a memory (430), and some of the illustrated components may be omitted or substituted. A system (400) according to one embodiment may be a server or a terminal. According to one embodiment, the processor (420) is a component capable of performing operations or data processing regarding the control and / or communication of each component of the system (400), and may be composed of one or more processors. The memory (430) may store information related to the method described above or store a program in which the method described above is implemented. The memory (430) may be volatile memory or non-volatile memory. The memory (430) may store various file data, and the stored file data may be updated according to the operation of the processor (120).

[0032] According to one embodiment, the processor (420) can execute a program and control the device (400). The code of the program executed by the processor (120) can be stored in memory (430). Operations of the processor (420) can be performed by loading instructions stored in memory (430).

[0033] FIG. 5 is a flowchart illustrating a method for evaluating the value of art assets using artificial intelligence according to one embodiment.

[0034] Although the process steps, methods, and algorithms shown in the flowchart of FIG. 5 have been described sequentially, they can be designed to operate in any appropriate order. That is, the steps of the process, method, and algorithm described in various embodiments of the present invention do not necessarily need to be performed in the order presented in the present invention. Furthermore, even if some steps are performed asynchronously, in other embodiments, these steps may proceed simultaneously.

[0035] In operation 510, the system (e.g., the system (400) of FIG. 4) can collect art-related data under the control of a processor (e.g., the processor (420) of FIG. 4). The system (400) can collect art-related data including artist information, artwork characteristic information, and market transaction information from multiple data sources.

[0036] The system (400) can collect data related to artworks from various data sources, such as global art platforms like Artsy and ArtNet, major gallery and auction house websites, and museum archives like the Guggenheim and the National Museum of Modern and Contemporary Art. The collected data may include artist information, such as the artist's exhibition history, career, and market evaluation information; artwork characteristic information, such as the size, medium, condition of preservation, and year of production of the artwork; and market transaction information, such as past transaction prices and auction prices of similar works, collected from Artprice, Sotheby's, Christie's, etc. The system (400) collects data in various ways, such as API integration and web scraping, and can integrate and normalize the collected data through an ETL pipeline.

[0037] An API (Application Programming Interface) is an interface that enables communication between different software systems, and in this invention, it can be utilized to automatically collect necessary information by accessing databases of major performance venues, art museums, music platforms, film distributors, etc. Web scraping is a technology that automatically extracts data from websites and can be used to collect information on irregular art events or review data from sites that do not provide APIs. An ETL (Extract, Transform, Load) pipeline can refer to a series of processes for extracting, transforming, and loading data.

[0038] In operation 520, the system (400) can train models using training data. The system (400) generates training data by normalizing collected data, and can train a current value prediction model, a price fluctuation prediction model, and an investment value evaluation model using the training data.

[0039] The system (400) can preprocess collected data and convert it into a form suitable for learning. The system (400) can train models such as Random Forest regression, XGBoost, LGBM, and linear regression to predict the current market value of the artwork. Additionally, it can train models to predict future price fluctuations using RNN, LSTM, GRU, Prophet, etc., and can also train classification models to evaluate the scarcity and long-term investment value of the artwork. During the training process of each model, performance can be evaluated through cross-validation, and hyperparameter optimization can be performed to improve the accuracy of the model.

[0040] Random Forest Regression is an ensemble learning method that derives predicted values ​​by constructing multiple decision trees, while XGBoost (eXtreme Gradient Boosting) and LGBM (Light Gradient Boosting Machine) are high-performance machine learning algorithms based on gradient boosting. Recurrent Neural Networks (RNNs) are neural network structures specialized for processing time-series data, and Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRUs) can be considered improved models that address the long-term dependency problem of RNNs. Prophet is a time-series forecasting library developed by Facebook that enables forecasting that automatically considers seasonality and trends. Cross-validation is a technique that divides data into multiple folds to evaluate a model's generalization performance, and hyperparameter optimization refers to the process of finding the optimal settings for a model through methods such as Grid Search, Random Search, and Bayesian Optimization.

[0041] In operation 530, the system (400) can calculate investment value using learned models. The system (400) can calculate the current market value, future price fluctuation prediction, and long-term investment value of a specific work of art using learned models, and can integrate the calculated results to generate a final evaluation result.

[0042] The system (400) can comprehensively evaluate the investment value of a specific artwork by utilizing learned models. It can calculate the current value and expected price range of the artwork through a current market value prediction model, and predict price fluctuation trends over the next five years through a price fluctuation prediction model. Additionally, it can calculate the scarcity index and long-term investment value index of the artwork through a scarcity evaluation model. The system (400) can generate a final investment value evaluation report by synthesizing these individual evaluation results, and in this process, it can also provide an interpretation of the evaluation results by utilizing explainable AI technologies such as SHAP and LIME.

[0043] SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) are techniques used to interpret and explain the prediction results of AI models, enabling the quantitative analysis and visualization of the contribution of each feature.

[0044] FIG. 6 is a flowchart illustrating a method for evaluating the value of art assets using artificial intelligence according to one embodiment.

[0045] In operation 610, a system (e.g., system (400) of FIG. 4) can extract data from artwork characteristic information, artist information, and market transaction information under the control of a processor (e.g., processor (420) of FIG. 4). The system (400) can extract physical characteristic data from the artwork characteristic information, including the size, medium, year of production, and preservation condition of the artwork. The system (400) can extract career data from the artist information, including the artist's exhibition history, awards history, major artwork sales history, and museum collection history. The system (400) can extract transaction data from the market transaction information, including the winning bid price, failure rate, number of auctions, and transaction volume of similar artworks over the past six months.

[0046] The system (400) can systematically extract data of various categories, such as the physical characteristics of the artwork, the artist's career details, and market transaction status, during the data extraction stage. In the artwork characteristic information, physical characteristics such as canvas size (e.g., No. 60, No. 100), materials used (e.g., oil paint, acrylic, watercolor), the time of production, and the condition of preservation can be extracted. In the artist information, career data such as exhibition history at major art museums, history of participation in international art fairs, art award winning history, and the status of collections at famous art museums can be extracted. In the market transaction information, market data such as auction data for the past six months (e.g., average winning bid of 500 million won), information on failed bids (e.g., failed bid rate of 15%), transaction frequency (e.g., average of 3 auction entries per month), and transaction scale can be extracted.

[0047] In operation 620, the system (400) can perform preprocessing on the extracted data. The system (400) can perform preprocessing on the extracted data by replacing missing values ​​with the mean value of the corresponding feature, quantifying categorical variables with one-hot encoding, and normalizing numerical variables with min-max scaling.

[0048] The system (400) can convert raw data extracted during the data preprocessing stage into a form suitable for machine learning models to learn. In the case of handling missing values, if the size of the artwork is missing, it can be replaced with the average size of other works by the artist, or if the preservation status is missing, an estimated value based on the year of production can be used. For categorical variables, for example, the medium of the artwork (e.g., oil painting, acrylic, watercolor, etc.) can be converted into binary vectors such as [1,0,0], [0,1,0], [0,0,1]. For numerical variables, values ​​such as the price or size of the artwork can be normalized to values ​​between 0 and 1 so that all variables have the same scale.

[0049] In operation 630, the system (400) can perform training of the current value prediction model using preprocessed data as input variables. The system (400) can set the number of decision trees to 100 and randomly select features equal to the square root of the total features when splitting each node. The system (400) can use a random forest algorithm to construct training data for each tree through bootstrap sampling and perform training of the current value prediction model using preprocessed data as input variables.

[0050] The system (400) can train a random forest-based current value prediction model using preprocessed data during the model training phase. The system (400) can generate 100 independent decision trees, and each tree can be trained by randomly selecting some of the total features (e.g., 5 if there are 25 total features). Each tree is trained with different training data, which is randomly sampled from the original data with duplicates allowed. For example, 7,000 of the total 10,000 data points are randomly selected to train each tree, and this process is repeated 100 times to build an ensemble model. The model trained in this way can predict the final value of a new work by averaging the prediction values ​​of the 100 trees.

[0051] FIG. 7 is a flowchart illustrating a method for evaluating the value of art assets using artificial intelligence according to one embodiment.

[0052] Although the process steps, methods, and algorithms shown in the flowchart of FIG. 7 have been described sequentially, they can be designed to operate in any appropriate order. That is, the steps of the process, method, and algorithm described in various embodiments of the present invention do not necessarily need to be performed in the order presented in the present invention. Furthermore, even if some steps are performed asynchronously, in other embodiments, these steps may proceed simultaneously.

[0053] In operation 710, the system (e.g., the system (400) of FIG. 4) can extract external event data under the control of a processor (e.g., the processor (420) of FIG. 4). The system (400) can extract external event data from market transaction information, including monthly average transaction price data for the last 36 months, monthly total transaction volume data, art market index data, major auction schedules, exhibition hosting, and changes in art market policies.

[0054] The system (400) can comprehensively collect macroscopic trends and major event information of the art market. From the market transaction information, market indicators such as transaction data over the past 36 months (e.g., average transaction price of 420 million won in January 2024, 450 million won in February, etc.), monthly transaction volume trends, and the K-ART index can be extracted. Additionally, event data such as major auction schedules, major exhibition information, and policy changes (e.g., implementation of the revised capital gains tax bill for artworks) can also be extracted.

[0055] In operation 720, the system (400) can perform preprocessing on the extracted time series data. The system (400) can perform preprocessing by applying a 3-sigma rule to the extracted time series data to remove outliers, decomposing and storing seasonal factors separately, and supplementing missing intervals using linear interpolation.

[0056] The system (400) can improve the quality of the extracted data and process it into a form suitable for time series analysis. For outlier processing, extreme transaction prices that deviate 3 standard deviations from the mean (e.g., a case where a work that was previously traded for 500 million won suddenly traded for 5 billion won) can be removed. For seasonal analysis, quarterly and monthly periodicity can be identified (e.g., a pattern where transactions increase in May and November of each year) and separated into separate components. For missing interval processing, the price of a period without transactions can be linearly estimated using the prices of the preceding and succeeding points.

[0057] In operation 730, the system (400) can train an LSTM network-based price fluctuation prediction model to predict price fluctuations over the next 24 months. The system (400) can configure the input sequence in 12-month increments and connect a first LSTM layer containing 64 hidden units and a second LSTM layer containing 32 hidden units in series. The system (400) can train an LSTM network-based price fluctuation prediction model to predict price fluctuations over the next 24 months by applying a Dropout rate of 0.2.

[0058] The LSTM (Long Short-Term Memory) network is a type of recurrent neural network capable of learning long-term patterns from time-series data, and in this invention, it can be utilized to predict price fluctuations of artworks. The input sequence refers to the length of the time-series data input into the neural network, and being structured in 12-month units means that future prices are predicted based on data from the previous year. A Hidden Unit is a basic unit that processes and stores information in the LSTM layer; the more units there are, the more complex patterns can be learned. The 64 Hidden Units in the first LSTM layer are responsible for initial feature extraction, while the 32 Hidden Units in the second LSTM layer further refine the extracted features and utilize them for final prediction. Dropout is a regularization technique designed to prevent overfitting in neural networks, and a ratio of 0.2 means that 20% of neurons are randomly deactivated during the training process. Through this structure, the system can predict art price fluctuations for the next 24 months, or two years, which can be used to provide investors with a medium-term investment outlook.

[0059] The system (400) can build a price fluctuation prediction model based on a recurrent neural network. The model uses 12-month time series data (e.g., monthly data from January to December 2022) as input, and can predict prices for the next 24 months. The first LSTM layer includes 64 neurons to learn long-term patterns of the time series, and the second LSTM layer includes 32 neurons to extract more abstract features. Dropout is applied to randomly deactivate neurons at a rate of 0.2 to prevent overfitting, which can operate by randomly turning off 20% of the total neurons during the training phase.

[0060] According to one embodiment, the system (400) can calculate a scarcity index by quantifying the total production quantity of the work, the existence of similar works in the market, the degree of originality of the production technique, and the preservation status of the work. In addition, a growth index can be calculated by combining the increase rate of the number of exhibitions during a specified period (e.g., the last 3 years), the increase rate of the sales price of the work, changes in the frequency of media exposure, and evaluation scores by art world experts.

[0061] The system (400) can calculate market awareness indicators by analyzing whether the artwork is in the collection of major art museums, the frequency of citation in the art market, the success rate of art auctions, and the frequency of online searches, and can calculate collector interest indicators by measuring the number of inquiries about purchasing artworks, the number of exhibition visitors, the number of social media related posts, and the number of online gallery views.

[0062] According to one embodiment, the system (400) can classify long-term investment value into five levels—best, high, medium, low, and worst—using the calculated indicators as input, by using a gradient boosting classifier that sets the maximum depth of the tree to 5, specifies the learning rate to 0.1, and sets the minimum number of sample splits to 20.

[0063] The system (400) can generate a reliability index for each model by calculating a confidence interval for the predicted value of the current market value, quantifying the time series prediction uncertainty of the price fluctuation prediction, and calculating the classification probability of the investment value evaluation. At this time, a weight of 40% is applied to the current market value prediction, 35% to the future price fluctuation prediction, and 25% to the long-term investment value evaluation, but the weight can be adjusted within a range of 5% according to the reliability index of each prediction to calculate an integrated evaluation score. The figures mentioned are merely examples and may vary depending on the settings.

[0064] According to one embodiment, the system (400) can calculate the confidence interval of the predicted value of the current market value, quantify the time series prediction uncertainty of the price fluctuation prediction, and calculate the classification probability of the investment value assessment to generate a reliability index for each model.

[0065] According to one embodiment, the system (400) may calculate an integrated evaluation score by applying a weight of 40% to the current market value prediction, a weight of 35% to the future price fluctuation prediction, and a weight of 25% to the long-term investment value evaluation, while adjusting the weights within a range of 5% according to the reliability index of each prediction. The figures mentioned are merely examples and may vary depending on the settings.

[0066] According to one embodiment, the system (400) can generate a final evaluation report by presenting an integrated evaluation score and grade, including individual evaluation results and reliability indices for each model, visualizing key evaluation grounds in charts and graphs, and analyzing investment risk factors and opportunity factors.

[0067] The system (400) can comprehensively organize and visually represent the evaluation results. The report presents the integrated evaluation score (e.g., 85.75 points) and investment grade (e.g., Grade A) on the main page, and can also display the detailed evaluation content of each model (e.g., current value 350 million won, expected value increase rate after 12 months 10%, long-term investment grade 'High') and reliability index (e.g., current value prediction reliability 90%, price fluctuation prediction reliability 85%, investment value evaluation reliability 80%). The basis for the evaluation can be visualized using time series graphs, bar graphs, etc. In addition, it can objectively analyze and present investment risk factors (e.g., possibility of increased supply of similar works, artist's reputation risk) and opportunity factors (e.g., scheduled solo exhibition at a major art museum, possibility of entering overseas markets).

[0068] FIG. 8 is a flowchart illustrating a method for evaluating the value of art assets using artificial intelligence according to one embodiment.

[0069] The system (400) may receive basic information, including the title of the work, the name of the author, the description of the work, and the type of the work, from an authenticated user and store it in memory under the control of the processor (420) in operation 810. The system (400) performs authentication through a username or phone number registered in a user authentication database and, if necessary, may additionally require two-step authentication, such as an SMS authentication code or fingerprint authentication.

[0070] In operation 820, the system (400) can perform an aesthetic evaluation through a four-dimensional evaluation mode including a color evaluation index, a visual effect index, a technical completeness index, and a spatial sense index, based on an image database built by preprocessing more than 500,000 domestic and international artwork images. The artwork images can be stored in the database after undergoing preprocessing processes such as resolution standardization, noise removal, and color correction.

[0071] In operation 830, the system (400) can apply a formula (e.g., Color = Ht1 + Vt2 + St3 + Ft4) that combines hue (H), value (V), purity (S), and richness (F) in color evaluation, perform texture analysis based on the LBP algorithm for visual effects, analyze image feature values ​​extracted through 3*3 convolution operations for technical completeness, and perform a ResNet-based 5-level grade evaluation for spatial sense to produce a final evaluation result.

[0072] In operation 840, the system (400) may generate an evaluation report by calculating a 5-level rating in four dimensions of theme, expression technique, composition, and modeling for evaluating the innovativeness of the work, and by applying a weighted combination formula (e.g., Match = (SSIM)ta + (CSIM)tb + (TSIM)tc + (PSIM)td) of composition similarity (SSIM), color similarity (CSIM), texture similarity (TSIM), and spatial similarity (PSIM) for evaluating the degree of similarity.

[0073] According to one embodiment, the system (400) can extract key feature points based on color, shape, and pattern from a plurality of two-dimensional or three-dimensional images taken from different angles. For example, features such as spiral patterns, contrast between dark blue and yellow, and rough brushstrokes can be identified in works such as Van Gogh's 'Starry Night', and these feature points can be efficiently matched using the SIFT (Scale-Invariant Feature Transform) algorithm and the KD-Tree data structure.

[0074] The system (400) can obtain 3D coordinates using the epipolar constraint relationship of matched feature points, generate a sparse point cloud from the obtained coordinates, and then expand it into a dense point cloud through a Multi-View Stereo algorithm. The system (400) can complete a 3D model by filling in the empty parts of the dense point cloud and performing texture mapping.

[0075] The system (400) can apply a ReLU activation function to each layer of a convolutional neural network (CNN) to normalize values ​​between 0 and 1, and compress a feature map through a pooling layer to extract key features. The system (400) inputs the extracted features into a DNN model composed of an input layer, a hidden layer, and an output layer to produce a similarity evaluation result between 0 and 1, and can determine that the closer the value is to 1, the higher the probability that it is genuine. In addition, the accuracy of the evaluation can be improved by applying the Structure from Motion (SFM) method to automatically restore camera motion and scene structure and generate a high-precision 3D point cloud.

[0076] According to one embodiment, the system (400) defines feature items including the year of production, type of work, subject of work, artist recognition, rarity of work, and production technique, and can perform quantification by sequentially assigning values ​​to each feature item based on the degree of proximity to the work to be evaluated. For example, a representative work of the early 20th century, such as Picasso's "Guernica," may be assigned normalized numerical values ​​such as 0.7 for the year of production, 0.95 for artist recognition, and 0.9 for rarity of work, and based on these quantified values, can generate first source domain sample feature data, first target domain sample feature data, and work to be evaluated feature data.

[0077] The system (400) can evaluate the similarity between domains by performing manifold learning based on an equal-distance mapping algorithm to calculate the similarity between the quantified value piA of the i-th feature item of a first source domain sample and the quantified value ciA of the i-th feature item of a work to be evaluated, and by setting the median value of D(p,c) as a similarity threshold. At this time, the system (400) can transfer first source domain samples with a calculated similarity smaller than the threshold to the target domain and display them as second target domain samples, and the samples that are not transferred can be displayed as second source domain samples.

[0078] The system (400) can construct a final training dataset including a first target domain sample, a second target domain sample, and a second source domain sample, apply a decision tree model based on the TrAdaBoost algorithm to the constructed training dataset, and initialize sample weights to generate a first sample weight vector. The system (400) can generate a transfer learning-based value evaluation model by performing iterative training once using feature data and value data of the weighted samples.

[0079] According to one embodiment, the system (400) can collect real-time transaction data and art auction data from the art market to construct a time-series database, and perform seasonal decomposition and trend analysis on the time-series data to extract market fluctuation patterns. The system (400) can collect and store transaction records of major auction houses such as Christie's and Sotheby's, gallery sales data, art fair transaction history, etc., and perform sentiment analysis through natural language processing on social media data and art-related news data to quantify changes in market preference for artists and artworks.

[0080] According to one embodiment, the system (400) can perform high-resolution image analysis of 4K or higher to evaluate the physical characteristics and artistic value of a work of art. It can analyze brushstroke characteristics, the condition of the canvas, color clarity, etc. through a convolutional neural network (CNN), and combine this with qualitative evaluations by experts such as curators and art historians to calculate a comprehensive quality index.

[0081] The system (400) trains a time series prediction model based on a recurrent neural network (RNN), such as LSTM or GRU, using the above market fluctuation pattern, sentiment analysis results, and quality index as inputs, and thereby can predict short-term (within 1 year), medium-term (1-3 years), and long-term (more than 3 years) value fluctuations of artwork. In addition, to evaluate the reliability of the predicted value fluctuations, it performs more than 10,000 Monte Carlo simulations, and based on the simulation results, calculates statistical indicators such as a 90% confidence interval to quantify the expected rate of return and investment risk.

[0082] According to one embodiment, the system (400) may utilize data from various asset markets to improve the accuracy of the valuation of artworks. The system (400) may collect data on stock price fluctuations of listed companies related to art and culture, data on real estate price fluctuations in regions where major galleries and art museums are located, trading data of bonds related to domestic and international cultural industries, and trading data of virtual assets related to digital art, such as NFTs.

[0083] The system (400) can calculate a correlation coefficient with the art market by performing time series analysis on the price fluctuation data of each collected asset group. The system (400) can construct an expanded training dataset by selectively utilizing only the data of asset groups for which the calculated correlation coefficient exceeds a preset threshold (e.g., 0.7) and combining it with existing art-related data. The system (400) can use this expanded dataset to retrain a current market value prediction model and a price fluctuation prediction model, and improve the accuracy of the final evaluation result by integrating the prediction values ​​of the retrained model and the prediction values ​​of the existing model in an ensemble manner.

[0084] According to one embodiment, the system (400) may utilize an ensemble learning method to precisely evaluate the value of a work of art. The system (400) may randomly assign an initial weight between 0 and 1 to each sample of a learning dataset and normalize it to generate a first weight vector whose sum is 1, then multiply each element of the first weight vector by the corresponding learning sample to generate a first weight processing sample, and store this by separating it into a feature data matrix and a value data vector. The system (400) may select an optimal splitting feature based on the information gain rate, but perform splitting only when the weight sum of the samples at each node is greater than or equal to a minimum threshold, and train a first decision tree model by limiting the maximum depth of the tree and the minimum number of leaf node samples, then use the trained model to calculate a predicted value for the first weight processing sample and calculate the squared mean of the difference between the predicted value and the actual value to calculate a first total error.

[0085] The system (400) can generate a second weight vector by applying an exponential function to the prediction error of each sample to calculate an error weight, multiplying it by an existing weight, and then normalizing it. Then, using the second weight vector, a second weight processing sample is generated, and a second decision tree model is trained in the same manner as the training process of the first decision tree model to calculate the second total error, and the ratio of the total error of the previous step to the current total error is calculated, and the next iteration can proceed only if the error ratio decreases.

[0086] The system (400) can perform iterations until a preset number of iterations l is reached or the error reduction rate becomes below a set threshold, thereby finally obtaining the first to first decision tree models and their respective total errors. The system (400) can calculate the reciprocal of the total error of each model and divide it by the sum of the reciprocals of the total errors to calculate the contribution coefficient of each model, and then linearly combine the first to first decision tree models using the contribution coefficient as a weight. Additionally, the system can apply an exponential moving average to the predicted values ​​of each model to improve prediction stability, calculate the prediction confidence interval of the final model to quantify uncertainty, extract feature data of the artwork to be evaluated, and perform normalization to calculate the basic evaluation value and confidence interval through the final value evaluation model.

Claims

1. In an art asset valuation system using artificial intelligence, Memory for storing instructions, Includes a processor, When the above instructions are executed by the processor, the system Collecting art-related data including author information, artwork characteristic information, and market transaction information from multiple data sources, and The collected data above is normalized to generate training data, and Using the above training data, a current value prediction model, a price fluctuation prediction model, and an investment valuation model are trained, and Using the above-mentioned learned models, the current market value, future price fluctuation predictions, and long-term investment value of a specific artwork are calculated, and An art asset valuation system that controls the integration of the above-mentioned results to generate a final evaluation result.

2. In Paragraph 1, When the above instructions are executed by the processor, the system Physical characteristic data including the size, medium, production year, and preservation condition of the artwork are extracted from the above artwork characteristic information, and Career data including the artist's exhibition history, awards, major work sales history, and museum collection history is extracted from the above artist information, and Transaction data including the winning bid price, failed bid rate, number of auctions, and transaction volume of similar works over the past six months is extracted from the above market transaction information, and Preprocessing is performed on the above extracted data by replacing missing values ​​with the mean value of the corresponding feature, quantifying categorical variables using one-hot encoding, and normalizing numerical variables using min-max scaling. Set the number of decision trees to 100, and when splitting each node, randomly select features equal to the square root of the total features, and An art asset valuation system that utilizes a random forest algorithm to construct training data for each tree through bootstrap sampling and controls the training of a current value prediction model using the aforementioned preprocessed data as input variables.

3. In Paragraph 1, When the above instructions are executed by the processor, the system Extracting external event data including monthly average transaction price data, monthly total transaction volume data, art market index data, major auction schedules, exhibition hosting, and changes in art market policies for the past 36 months from the above market transaction information, and Preprocessing is performed on the extracted time series data by applying the 3-sigma rule to remove outliers, decomposing and separately storing seasonal factors, and supplementing missing intervals using linear interpolation, The input sequence is configured in 12-month units, and a first LSTM layer containing 64 hidden units and a second LSTM layer containing 32 hidden units are connected in series, An art asset valuation system that controls the training of an LSTM network-based price fluctuation prediction model to predict future price fluctuations over the next 24 months by applying a dropout rate of 0.

2.

4. In Paragraph 1, When the above instructions are executed by the processor, the system Calculate a scarcity index by quantifying the total production quantity of the work, the existence of similar works in the market, the degree of originality of the production technique, and the preservation status of the work, and A growth indicator is calculated by synthesizing the growth rate of the number of exhibitions during a designated period, the rate of increase in artwork sales prices, changes in the frequency of media exposure, and evaluation scores from art world experts, and Calculate market awareness indicators by analyzing whether the work is held in major art museums, frequency of citation in the art market, art auction success rates, and online search frequency, and It calculates collector interest indicators by measuring the number of artwork purchase inquiries, the number of exhibition visitors, the number of related social media posts, and online gallery views, An art asset valuation system that uses a gradient boosting classifier with a maximum tree depth set to 5, a learning rate set to 0.1, and a minimum number of sample splits set to 20, and controls the classification of long-term investment values ​​into 5 stages—best, high, medium, low, and lowest—using the calculated indicators as input.

5. In Paragraph 1, When the above instructions are executed by the processor, the system Additionally, collect stock price fluctuation data of art or culture-related companies in the stock market, real estate price fluctuation data of regions where galleries and museums are located in the real estate market, bond trading data related to domestic and international cultural industries, and virtual asset trading data related to digital art, and Time series analysis is performed on each of the above collected data to calculate the correlation coefficient with the art market, and Select price fluctuation data for asset groups where the correlation coefficient calculated above exceeds a set threshold, and By combining selected price fluctuation data with existing art-related data to construct an extended training dataset, Using the aforementioned extended training dataset, the current market value prediction model and the price fluctuation prediction model are retrained, and By integrating the predicted values ​​produced by retrained models and the predicted values ​​of existing models using ensemble techniques, the accuracy of the final evaluation result is improved, Calculate the confidence interval for the predicted value of the current market value, quantify the time series forecast uncertainty of price fluctuation predictions, and calculate the classification probability of investment valuation to generate a reliability index for each model, and A weight of 40% is applied to current market value predictions, 35% to future price fluctuation predictions, and 25% to long-term investment value assessments, but the weights are adjusted within a 5% range according to the reliability index of each prediction to calculate an integrated evaluation score. An art asset valuation system that controls the generation of a final evaluation report by presenting integrated evaluation scores and grades, including individual evaluation results and reliability indices for each model, visualizing key evaluation grounds in charts and graphs, and analyzing investment risk factors and opportunity factors.

6. In Paragraph 1, When the above instructions are executed by the processor, the system Authentication is performed through a username or phone number registered in a user authentication database, and basic information including the work title, author name, work description, and work type is received from the authenticated user and stored in the memory. Based on an image database constructed by preprocessing over 500,000 domestic and international artwork images, aesthetic evaluation is performed through a 4D evaluation mode including a color evaluation index, visual effect index, technical completeness index, and spatial perception index, and In color evaluation, a formula combining weighted hue (H), value (V), purity (S), and richness (F) is applied; for visual effects, texture analysis based on the LBP algorithm is performed; for technical completeness, image feature values ​​extracted through 3x3 convolution operations are analyzed; and for spatial perception, a ResNet-based 5-level rating evaluation is performed to produce the final evaluation result. To evaluate the innovativeness of the work, a 5-level rating is calculated across four dimensions of theme, expression technique, composition, and modeling, and a formula combining composition similarity (SSIM), color similarity (CSIM), texture similarity (TSIM), and spatial similarity (PSIM) is applied, and An art asset valuation system that controls the generation of evaluation reports based on a detailed evaluation description database and a comprehensive evaluation and growth proposal database.

7. In Paragraph 1, When the above instructions are executed by the processor, the system Extracting key feature points based on color, shape, and pattern from multiple 2D images captured from different angles, and Feature points are matched using the KD-TREE method with block-based matching and SIFT features, and 3D coordinates are obtained using the epipolar constraint relationships of matched feature points, a sparse point cloud is generated from the obtained coordinates, and then expanded to generate a dense point cloud, and The 3D model is completed by filling in the blank parts of the above dense point cloud and performing texture mapping, and The ReLU activation function is applied after each convolution layer to normalize values ​​between 0 and 1, and feature maps are compressed and key features are extracted through pooling layers, and The extracted features are input into a DNN model composed of an input layer, a hidden layer, and an output layer to produce a similarity evaluation result between 0 and 1, wherein a value closer to 1 is determined to be more likely to be genuine. An art asset valuation system that applies the Structure from Motion (SFM) method to automatically restore camera motion and scene structure and generates a high-precision 3D point cloud to improve the accuracy of the evaluation.

8. In Paragraph 1, When the above instructions are executed by the processor, the system Define characteristic items including the year of production, type of work, subject matter, artist recognition, rarity, and production technique, and After performing quantification by sequentially assigning values ​​to each feature item based on proximity to the work to be evaluated, the first source domain sample feature data, the first target domain sample feature data, and the work to be evaluated feature data are generated based on the quantified values. Manifold learning based on an equidistant mapping algorithm is performed to calculate the similarity between the quantified value piA of the i-th feature item of the first source domain sample and the quantified value ciA of the i-th feature item of the work to be evaluated, and the similarity between domains is evaluated by setting the median value of D(p,c) as the similarity threshold, and First source domain samples with a similarity smaller than the threshold value are transferred to the target domain and marked as second target domain samples, After marking the samples that were not transferred as second source domain samples, Constructing a final training dataset including a first target domain sample, a second target domain sample, and a second source domain sample, An art asset valuation system that controls the generation of a transfer learning-based valuation model by applying a decision tree model based on the TrAdaBoost algorithm to the above-configured training dataset, initializing sample weights to generate a first sample weight vector, and performing iterative learning once using feature data and value data of the weighted samples.

9. In Paragraph 1, When the above instructions are executed by the processor, the system A time-series database is constructed by collecting real-time transaction data and art auction data from the art market, and market fluctuation patterns are extracted by performing seasonal decomposition and trend analysis on the said time-series data. By collecting social media data and art-related news data, performing sentiment analysis on artists and artworks through natural language processing, and quantifying changes in artist recognition and artwork market preference, High-resolution image analysis is performed to evaluate the physical characteristics, artistic value, and preservation status of artworks, while extracting characteristics of the artwork's style, technique, and material usage through a Convolutional Neural Network (CNN), and combining this with expert evaluation data to calculate the artwork's quality index. A recurrent neural network (RNN)-based time series forecasting model is trained using the above market fluctuation patterns, sentiment analysis results, and quality index as inputs, and through this, short-term, medium-term, and long-term value fluctuations of artworks are predicted. An art asset valuation system that performs Monte Carlo simulation to evaluate the reliability of predicted value fluctuations and controls the quantification of investment risk by calculating a prediction interval based on the simulation results.

10. In Paragraph 1, When the above instructions are executed by the processor, the system For each sample of the training dataset, an initial weight between 0 and 1 is randomly assigned and normalized to generate a first weight vector whose sum is 1, and then each element of the first weight vector is multiplied by the corresponding training sample to generate a first weight processing sample, and this is stored by separating it into a feature data matrix and a value data vector. After training a first decision tree model by selecting an optimal splitting feature based on the information gain rate, performing splitting only when the weighted sum of samples at each node is greater than or equal to a minimum threshold, and limiting the maximum depth of the tree and the minimum number of leaf node samples, the trained model is used to calculate a predicted value for the first weighted processing sample, and the first total error is calculated by calculating the mean square of the difference between the predicted value and the actual value. Error weights are calculated by applying an exponential function to the prediction error of each sample, and a second weight vector is generated by multiplying these by the existing weights and normalizing them. After generating a second weight processing sample using the above second weight vector, the second decision tree model is trained in the same manner as the training process of the first decision tree model to calculate the second total error, and Calculate the ratio of the total error from the previous step to the current total error, and proceed to the next iteration only if the error ratio decreases, while updating weights and training a new model in the same way as the previous step in each iteration, Iterations are performed until a preset number of iterations l is reached or the error reduction rate becomes below a set threshold, thereby finally obtaining the decision tree models from the first to the lth and their respective total errors, and Calculate the reciprocal of the total error of each model and divide it by the sum of the reciprocals of the total errors to calculate the contribution coefficient of each model, and then linearly combine the decision tree models from the first to the first using the said contribution coefficient as weights. Improve prediction stability by applying an exponential moving average to the predicted values ​​of each model, and quantify uncertainty by calculating the prediction confidence interval of the final model, and An art asset valuation system that extracts characteristic data of an artwork to be evaluated, performs normalization, and then controls the calculation of a basic valuation value and a confidence interval through a final valuation model.

Images