Ai model-based unique painting style information generation and artist inference method

Abstract

Disclosed is an AI model-based painting style information generation and artist inference method. The AI model-based painting style information generation and artist inference method comprises the steps of: extracting a feature map of an input image by means of a backbone network; generating a gram matrix by calculating an inner product of the extracted feature map; generating a unique style feature vector for an artist of the input image by performing representation learning on the gram matrix; and performing classification learning inferring a specific artist from the unique style feature vector. This method may provide a highly reliable imitation determination and estimation of an artist with respect to the style or painting style of an image by means of machine learning.

Description

AI Model-Based Method for Generating Unique Art Style Information and Inferring Author

[0001] The present invention relates to a method for generating unique art style information and inferring an artist based on an AI model. Specifically, it relates to a method for generating unique art style information and inferring an artist based on an AI model that can objectively analyze the similarity of imitation works through AI by generating feature vectors of an artist's art style through machine learning and performing machine learning to classify the artist therefrom.

[0002] This invention is a study conducted by the Korea Electronics Technology Institute under the project title "Development of Content Creation and Sharing Platform Technology Equipped with a Generative AI Model Blocking Malicious Use" for the "Cultural Technology Research and Development" project, with support from the Ministry of Culture, Sports and Tourism (Project No. 00441174, Project Unique No. 2370000125).

[0003] Neural Network Style Transfer is a computer vision technique that creates a new image by combining the content of one image with the style of another. This technique applies the visual style of another image while preserving the content of the original image.

[0004] FIG. 10 illustrates an example of style transfer in which the styles of other famous paintings are applied to the first original photograph. Through style transfer, the content of the original photograph—namely, the houses arranged diagonally across the riverbank—is maintained, while the styles or artistic styles of William Turner’s *The Shipwreck of the Minotaur*, Pablo Picasso’s *Untitled*, Vincent van Gogh’s *Starry Night*, Edvard Munch’s *The Scream*, and Wassily Kandinsky’s *Composition VII* are fused.

[0005] Style transfer is trained using a total loss function that combines content loss and style loss via deep neural network learning techniques. If the original content is strongly preserved, the intended art style is not clearly visible, while applying a strong art style leads to significant loss of the original content.

[0006] With the advancement of artificial intelligence, various forms of generative content are flooding the market, leading to instances where artists' intellectual property (IP) is infringed upon due to artistic styles similar to those of the works used for training. However, due to the expansion of digital distribution, it is currently difficult to detect and classify copyrights for all works, and even when detected, establishing clear criteria for distinguishing between imitation and plagiarism remains challenging.

[0007] Traditional copyright infringement refers to the temporary or permanent use of relevant content through methods such as photography, filming, copying, printing, recording, video recording, or reproduction, and copyright infringement can become an even greater problem with the development of generative AI.

[0008] As such technological advancements make it easy to produce and distribute imitation works, consistent and objective standards and methods are needed to judge and verify the style of images in order to protect the rights of genuine artists and minimize damages.

[0009] The present invention originates from such recognition, and the objective of the present invention is to provide a method for generating unique art style information and inferring an artist based on an AI model, which can objectively analyze the similarity of imitation works through AI by generating feature vectors of the artist's art style through machine learning and executing machine learning to classify the artist therefrom.

[0010] A method for generating art style information and inferring an artist based on an AI model according to an embodiment of the present invention for achieving such objectives comprises: a step of extracting a feature map of an input image by a backbone network; a step of generating a Gram matrix by calculating the inner product of the extracted feature maps; a step of generating an identified style feature vector for the artist of the input image by performing representation learning on the Gram matrix; and a step of performing classification learning to infer a specific artist from the identified style feature vector.

[0011] Here, the backbone network may be a pre-trained model to extract feature information of the shape and color of an image from a predetermined dataset.

[0012] In this case, the pre-trained model may be any one of VGGNet, ResNet, AlexNet, DenseNet, and EfficientNet.

[0013] Meanwhile, the above Gram matrix includes correlations between the extracted feature maps, and the step of performing the representation learning can perform representation learning to generate the unique style feature vector based on the correlations included in the Gram matrix.

[0014] Meanwhile, the step of performing the above-mentioned expression learning can be performed using a loss function to distinguish unique style feature vectors generated from drawings by different artists.

[0015] Here, the step of performing the classification learning can be performed using a loss function for estimating the author from the unique style feature vector obtained from the representation layer where the representation learning was performed.

[0016] Meanwhile, a computer-readable recording medium recording a program according to another embodiment of the present invention may be used to execute the above-described method for generating art style information based on an AI model and inferring an artist on a computer.

[0017] Furthermore, an AI model-based art style information generation and author inference system according to another embodiment of the present invention includes an input unit that receives an image and a processor that extracts a feature map of an input image using a backbone network, generates a Gram matrix by calculating the inner product of the extracted feature maps, generates an identified style feature vector for the author of the input image by performing representation learning on the Gram matrix, generates an AI model by performing classification learning to infer a specific author from the identified style feature vector, and performs an operation to estimate the author of a target image using the generated AI model.

[0018] Here, the backbone network may be a pre-trained model to extract feature information of the shape and color of an image from a predetermined dataset.

[0019] In this case, the above-mentioned pre-trained model may be any one of VGGNet, ResNet, AlexNet, DenseNet, and EfficientNet.

[0020] Additionally, the Gram matrix includes correlations between the extracted feature maps, and the processor can perform representation learning to generate the unique style feature vector based on the correlations included in the Gram matrix.

[0021] Here, the processor can perform the representation learning using a loss function to distinguish unique style feature vectors generated from drawings by different artists.

[0022] In this case, the processor can perform the classification learning using a loss function for estimating the author from the unique style feature vector obtained from the representation layer that performed the representation learning.

[0023] The method for generating unique art style information based on an AI model and inferring the artist according to one embodiment of the present invention as described above can provide a basis for determining whether an art style in an artistic domain has been imitated or infringed upon using objective digital data by providing a promised objective AI learning algorithm and criteria for distinguishing artist styles.

[0024] FIG. 1 is a block diagram showing the configuration of an AI model-based art style information generation and author inference system according to an embodiment of the present invention;

[0025] FIG. 2 is a block diagram showing the configuration of an AI model-based art style information generation and author inference algorithm according to an embodiment of the present invention;

[0026] FIG. 3 is a diagram schematically showing the configuration of the algorithm of FIG. 2;

[0027] FIG. 4 is a diagram illustrating the configuration of the backbone layer;

[0028] FIG. 5 is a diagram for explaining Gram matrix operations;

[0029] FIG. 6 is a diagram illustrating the presentation layer and the classification layer;

[0030] FIG. 7 is a block diagram showing the specific configuration of an AI model-based art style information generation and author inference system according to an embodiment of the present invention;

[0031] FIG. 8 is a diagram schematically showing a software module stored in the storage unit of the AI ​​model-based art style information generation and artist inference system of FIG. 7;

[0032] FIG. 9 is a flowchart illustrating a method for generating art style information and inferring the artist based on an AI model; and,

[0033] FIG. 10 is a diagram illustrating an example of a conventional neural network style transfer technology.

[0034] The present invention will be described in more detail below with reference to the drawings. Furthermore, in describing the present invention, detailed descriptions of related known functions or configurations are omitted if it is determined that such detailed descriptions would unnecessarily obscure the essence of the invention. Additionally, the terms described below are defined considering their functions in the present invention, and these may vary depending on the intentions or relationships of the user or operator. Therefore, their definitions should be based on the content throughout this specification.

[0035]

[0036] FIG. 1 is a block diagram showing the configuration of an AI model-based art style information generation and author inference system according to one embodiment of the present invention.

[0037] Referring to FIG. 1, an AI model-based art style information generation and artist inference system (100) according to one embodiment of the present invention includes a processor (110), an input unit (120), a communication unit (130), a display (140), and a storage unit (150). The AI ​​model-based art style information generation and artist inference system (100) may be implemented as a general-purpose or special-purpose computing system (100).

[0038] The processor (110) uses an AI model to generate feature information representing the artist's style and executes an algorithm to infer which artist the style reflected in the image belongs to based on the features. The processor (110) basically performs operations to extract features of a painting style based on image analysis AI and can execute learning to classify from the artist's unique style information whether the composition, color combination, drawing style, etc., were imitated by another person even if the painting or photograph of a specific artist was not copied exactly, or whether it is an unknown replica (authentic) of the true artist. The specific algorithm will be described later with reference to FIGS. 2 to 6.

[0039] The input unit (120) receives an image. The image may be image data of a drawing or photographic work. The input unit (120) may be a camera that captures a physical work, and may be a driver connector, a communication port, or an interface for connecting a device to the system (100).

[0040] The communication unit (130) is configured to establish communication with other devices. The communication unit (130) may include a driver that transmits and receives signals according to a well-known wired or wireless communication protocol. The communication unit (130) can collect information on the artist and artwork requiring learning online, connect with external devices requesting image analysis, such as user terminals and servers, and transmit the analyzed results.

[0041] The display (140) can output a visual screen. The display (140) can output a user interface screen for the operation and execution of the system (100), and can output the process and results of an algorithm for generating art style information and inferring the artist based on the AI ​​model.

[0042] The storage unit (150) may store basic software programs for operating the system (100), software modules for AI model-based style information generation and artist inference algorithms, and related data. The storage unit (150) is not limited to the built-in drive of the system (100) and may include a network-connected cloud storage.

[0043] The following describes an AI model-based art style information generation and author inference algorithm according to an embodiment of the present invention, which is executed by a processor (110).

[0044]

[0045] FIG. 2 is a block diagram showing the configuration of an AI model-based art style information generation and author inference algorithm according to one embodiment of the present invention.

[0046] First, an image (IMG) is input through the input unit (120). Initially, the image (IMG) is an authentic work reflecting the artist's characteristic painting style. Analysis by an AI model of a specific artist's painting or photograph is performed first, followed by inference regarding imitations or derivative works by others.

[0047] Next, an image is input into the backbone (210) and a feature map of the image is extracted. The backbone (210) network is also called a feature extracting network, and a feature map representing the image is generated through layer operations of the neural network using a deep learning-based algorithm. The backbone (210) is a representative AI model that is pre-trained with a large dataset and can generate a feature map of an image, and can be any one of VGGNet, ResNet, AlexNet, DenseNet, and EfficientNet.

[0048] The Gram matrix (GRAM MTX) (220) is the inner product of the feature maps and is calculated as follows.

[0049]

[0050] Here, V is the feature map and V T is the transpose of feature map V.

[0051] The Gram matrix of the feature map calculated in this way consists of values ​​that reflect the correlation between the channels constituting the feature map extracted by the backbone.

[0052] The representation layer (230) is a layer that represents a Gram matrix processed from a specific image as a mutually classifiable phenotype, and is composed of representation learning (or feature learning). The feature vector of the Gram matrix that has passed through the representation layer (230) enables it to be linearly well distinguished in the classification space.

[0053] The classification layer (240) is a layer that classifies images into specific authors (AT) from the gram matrix of the feature map. The classification layer (240) can be constructed through classification learning that corresponds actual authors with style feature vectors unique to those authors.

[0054] The result of the algorithm may be a sorting of multiple classifiable authors in order of probability (logit). That is, the inference of the author is performed by the algorithm of the present invention.

[0055]

[0056] Figure 3 is a diagram schematically showing the configuration of the algorithm of Figure 2.

[0057] Referring to FIG. 3, the backbone network (210) can be composed of VGGNet (16 weight layers). VGG passes through CNN-based convolutional layers and fully connected layers (FC layers) to output vectors that can be used for various applications, such as object recognition, in the subsequent application layer. However, in the present invention, the feature map output during pooling in the convolutional layer of VGG is utilized.

[0058] As feature maps undergo repeated convolution (convolution + ReLU), pixel-level information is lost, but it is expected that feature maps with discriminative power will be extracted with a reduced batch size. However, the number of layers and performance are not always proportional. For example, feature maps are extracted from a 5-layer pooling layer.

[0059] Pooling typically maintains the max pooling used in the pre-training of the existing VGG model, or average pooling may be applied through redesign.

[0060] The next Gram matrix layer (220) unfolds a 3-dimensional feature map with width (w) * height (h) * channel (c) into 2 dimensions and calculates the inner product.

[0061] The representation layer (230) is a layer that is trained to automatically extract features of an input value through representation learning (also called feature learning). Representation learning is a type of machine learning in which a machine learns to form concepts on its own regarding an input image or an object contained in an image, and to understand them as features that distinguish it from other objects, without relying on human understanding.

[0062] The output that passes through the presentation layer (230) is a feature vector that can distinguish the input image from other images. And, if the input image is an authentic work by that artist, this feature vector is named an identified style feature vector that can distinguish the artist.

[0063] The following classification layer (240) is a layer that matches the feature vector, which is uniquely located in a position distinguishing it from the works of other artists in space (h) after passing through the representation layer (230), with the artist of the actual input image. The result of estimating the artist may be the probability (logits) of belonging to each artist belonging to the artist pool of the dataset.

[0064]

[0065] Figure 4 is a diagram illustrating the configuration of the backbone layer.

[0066] A backbone layer (210) according to one embodiment of the present invention is composed of a pre-trained model of VGGNet. One convolutional layer (Conv) is composed of a convolution operation (e.g., 3x3) and a ReLU function (activation function), and a feature map (FM#) is extracted by passing through a pooling layer before entering the next convolutional layer (Conv). As an example, the present invention uses the FM5 feature map of the last Conv5 layer.

[0067]

[0068] Figure 5 is a diagram illustrating Gram matrix operations.

[0069] The FM5 feature map has a batch size of 7x7x512 = 25,088. The 3D feature map is unfolded into 2D (Fig. 5a), and the Gram matrix is ​​obtained by performing the inner product (Equation 1) (Fig. 5b). The Gram matrix relaxes the channel-specific independence of the feature map and allows the style revealed by the correlation between all channel layers to be expressed.

[0070]

[0071] Figure 6 is a diagram illustrating the presentation layer and the classification layer.

[0072] The representation layer (230) is formed by representation learning or feature learning. The feature learning model used for feature learning may be a CNN-based model that is effective for image analysis. For learning, well-known learning methods such as supervised learning, unsupervised learning, or self-supervised learning may be used to output a vector capable of sufficiently distinguishing images of different painting styles, and an appropriate loss function may be designed for this purpose.

[0073] The feature representation vectors that have passed through the representation layer are input into a classification layer (240) for a specific purpose application, in the case of the present invention, for classifying the author of the target image. The classification layer is generated by performing a well-known classification learning. Style feature vectors unique to a specific author's work are labeled as that specific author, and supervised learning can be performed to minimize the loss function through a loss function that compares with the author predicted from the result of the classification layer. In this case, if there are multiple works by one author, classification learning can be performed so that all of those multiple works are classified as belonging to that author.

[0074]

[0075] Subsequently, in a specific embodiment of the present invention, when determining imitation or infringement regarding a similar drawing or photograph used without authorization and identifying the rights holder, the image subject to judgment is input into the AI ​​model-based style information generation and artist inference module, which outputs a result estimating it to belong to a specific artist based on a certain probability. If the distance between the estimated style feature vector and the unique style feature vector is greater than a certain level or the probability of the estimated artist is lower than a certain level, it may be considered an original work; however, if it is classified as a work by a specific artist with a high probability, it may be determined that the work imitates that artist's work, and that using it without authorization constitutes an infringement of that artist's copyright, etc.

[0076]

[0077] FIG. 7 is a block diagram showing the specific configuration of an AI model-based art style information generation and author inference system according to one embodiment of the present invention.

[0078] Referring to FIG. 7, the system (100) includes a processor (110), an input unit (120), a communication unit (130), a display (140), and a storage unit (150). The general description of the input unit (120), the communication unit (130), the display (140), and the storage unit (150) is based on the description with reference to FIG. 1.

[0079] The processor (110) controls the components of the system (100) overall.

[0080] The processor (110) includes RAM (111), ROM (112), main CPU (113), graphics processing unit (114), first to n interfaces (115-1 to 115-n), and a bus (116).

[0081] RAM (111), ROM (112), main CPU (113), graphics processing unit (114), first to n interfaces (115-1 to 115-n), etc. can be connected to each other via a bus (116).

[0082] The first to n interfaces (115-1 to 115-n) are connected to various components (120, 130, 140, 150). One of the interfaces may be a network interface that is connected to an external device via a network. For example, the interface (115) may be connected to another Bluetooth-enabled device via short-range wireless communication such as Bluetooth.

[0083] The main CPU (113) accesses the storage unit (150) and performs booting using the O / S stored in the storage unit (150). Then, it performs various operations using various programs, content, data, etc. stored in the storage unit (150).

[0084] In particular, the main CPU (113) can perform a series of operations to classify a specific author by extracting a feature map from a target image input to the input unit (120) using a backbone model, calculating a Gram matrix, and generating a feature representation vector.

[0085] A set of instructions for booting the system is stored in the ROM (112). When a turn-on command is input and power is supplied, the main CPU (113) copies the O / S stored in the storage unit (150) to the RAM (111) according to the instructions stored in the ROM (112), and executes the O / S to boot the system. When booting is complete, the main CPU (113) copies various application programs stored in the storage unit (150) to the RAM (111), and executes the application programs copied to the RAM (111) to perform various operations.

[0086] The graphics processing unit (114) generates a screen containing various objects such as icons, images, and text using a calculation unit (not shown) and a rendering unit (not shown). The calculation unit (not shown) calculates attribute values ​​such as coordinate values, shape, size, and color for each object to be displayed according to the layout of the screen based on a received control command. The rendering unit (not shown) generates a screen with various layouts containing objects based on the attribute values ​​calculated by the calculation unit (not shown).

[0087] In particular, the graphics processing unit (114) can implement objects generated by the main CPU (113) as a GUI (Graphic User Interface), icon, user interface screen, etc. The graphics processing unit (114) can configure a screen that displays an input image, an estimated artist, an estimated probability, a style feature vector of the input image, and a work (e.g., backward processing) corresponding to the artist's unique style feature vector of the recent distance (cluster distribution) within the representation space (h).

[0088] The display unit (160) can display a screen generated by the graphics processing unit (114).

[0089] The storage unit (150) stores various data, such as firmware or O / S (Operating System) software modules and various multimedia content, for the operation of the AI ​​model-based art style information generation and artist inference system (100).

[0090] In particular, the storage unit (150) can store an algorithm for performing the method, a dataset for learning, a generated AI model, and an input image.

[0091]

[0092] Figure 8 is a schematic diagram showing a software module stored in the storage unit of the AI ​​model-based art style information generation and author inference system of Figure 7.

[0093] Referring to FIG. 8, the storage unit (150) includes a backbone network module (151), a Gram matrix calculation module (152), a representation learning module (153), and a classification learning module (154).

[0094] The backbone network module (151) is executed by the processor (110) to enable the extraction of feature maps from an input image using a pre-trained backbone model. The backbone network module (151) may provide a CNN-based backbone model, and the feature map may be a feature map pooled in the deepest convolutional layer among several convolutional layers.

[0095] The Gram matrix calculation module (152) can be executed by the processor (110) to provide a calculation that unfolds the feature map by channel, converts it into a 2D matrix, and then performs an inner product.

[0096] The representation layer module (153) is executed by the processor (110) and can provide CNN-based representation learning, a loss function for this, and a representation layer model learned therefrom, which learns the features of the input values ​​(gram matrix) and outputs a feature vector or a unique style feature vector that can be distinguished from the style of other images, without the designer having to design the distinguishing features of the input images one by one.

[0097] The classification layer module (154) can provide classification learning that classifies (infers) unique style feature vectors distinguished from one another by representation learning as the author of the work (feature vector), a loss function for this, and a classification layer model learned accordingly. The classification model may be fine-tuned to be specialized for the task of classifying (calculating estimated probabilities) the authors of the author pool from a pre-trained model such as an existing CNN-based VGG.

[0098]

[0099] Figure 9 is a flowchart illustrating a method for generating art style information and inferring the artist based on an AI model.

[0100] Referring to Fig. 9, the method for generating art style information and inferring the artist based on an AI model can start by inputting a dataset containing images of the artist's works to be learned (S710).

[0101] First, the method for generating art style information and inferring the artist based on an AI model includes a step (S920) of extracting a feature map of an input image by a backbone network. The backbone network is a model pre-trained to extract feature information of the shape and color of an image from a predetermined dataset, and may be any one of VGGNet, ResNet, AlexNet, DenseNet, and EfficientNet.

[0102] Next, the method for generating art style information and inferring the artist based on an AI model includes the step (S930) of generating a Gram matrix by calculating the inner product of the extracted feature maps. The Gram matrix is ​​composed of values ​​that reflect the correlation between the extracted feature maps and are calculated according to Equation 1.

[0103] Furthermore, the method for generating art style information and inferring the artist based on an AI model includes a step (S940) of generating an identified style feature vector for the artist of an input image by performing representation learning on a Gram matrix. The representation learning can be performed by generating the identified style feature vector based on the correlations contained in the Gram matrix and using a loss function to distinguish the identified style feature vectors generated from paintings by different artists.

[0104] In addition, the method for generating art style information and inferring an artist based on an AI model includes a step (S950) of performing classification learning to infer a specific artist from a unique style feature vector. The classification learning can be performed using a loss function to estimate the artist from the unique style feature vector obtained from the expression layer where expression learning was performed.

[0105] The above steps (S920 to S950) can be performed periodically to learn new artists and works. Subsequently, when a target image is input for which an artist is to be inferred from its style, the AI ​​model-based style information generation and artist inference method can extract its feature map, obtain a style feature vector by the representation model of the representation layer learned from the gram matrix of the feature map, and output an artist inferred from that feature vector by the classification model of the classification layer.

[0106] The AI ​​model-based art style information generation and author inference method can be terminated by the output of the author with the highest probability or authors by probability, or by user command.

[0107]

[0108] Meanwhile, a non-transitory computer-readable medium may be provided that stores a program for sequentially performing the method for generating art style information and inferring the artist based on an AI model according to the present invention.

[0109] A non-transient readable medium refers to a medium that stores data semi-permanently and can be read by a device, rather than a medium that stores data for a short period of time, such as a register, cache, or memory. Specifically, the various applications or programs described above may be stored and provided on non-transient readable media such as CDs, DVDs, hard disks, Blu-ray discs, USBs, memory cards, and ROMs.

[0110] Furthermore, although preferred embodiments of the present invention have been illustrated and described above, the present invention is not limited to the specific embodiments described above. It is understood that various modifications can be made by those skilled in the art without departing from the essence of the invention as claimed in the claims, and such modifications should not be understood individually from the technical spirit or perspective of the present invention.

Claims

1. In the method for generating art style information and inferring the author based on an AI model, A step of extracting a feature map of an input image by a backbone network; A step of generating a Gram matrix by calculating the inner product of the extracted feature maps; A step of generating an identified style feature vector for the author of the input image by performing representation learning on the above Gram matrix; and A method for generating art style information and inferring an author based on an AI model, comprising the step of performing classification learning to infer a specific author from the above-mentioned unique style feature vector.

2. In Paragraph 1, The above-mentioned backbone network is an AI model-based method for generating art style information and inferring the artist, which is a model pre-trained to extract feature information of the shape and color of an image from a predetermined dataset.

3. In Paragraph 2, The above-mentioned pre-trained model is any one of VGGNet, ResNet, AlexNet, DenseNet, and EfficientNet, an AI model-based method for generating art style information and inferring the artist.

4. In Paragraph 1, The above Gram matrix includes correlations between the extracted feature maps, and The step of performing the above expression learning is, AI model-based method for generating art style information and inferring author, which performs expression learning to generate the unique style feature vector based on the correlations included in the above Gram matrix.

5. In Paragraph 1, The step of performing the above expression learning is, A method for generating art style information and inferring artists based on an AI model, wherein the above-mentioned expression learning is performed using a loss function to distinguish unique style feature vectors generated from paintings by different artists.

6. In Paragraph 5, The step of performing the above classification learning is, AI model-based method for generating art style information and inferring author, wherein the classification learning is performed using a loss function for estimating the author from a unique style feature vector obtained from an expression layer in which the expression learning is performed.

7. A computer-readable recording medium having a program stored on it for executing on a computer the method for generating art style information and inferring the author based on an AI model as described in any one of paragraphs 1 through 6.

8. In an AI model-based art style information generation and author inference system, Input unit for receiving an image; and The feature map of the input image is extracted by a backbone network, and A Gram matrix is ​​generated by calculating the inner product of the above extracted feature maps, and Representation learning is performed on the above Gram matrix to generate an identified style feature vector for the artist of the above input image, and An AI model is created by performing classification learning to infer a specific author from the above-mentioned unique style feature vector, and An AI model-based art style information generation and author inference system comprising: a processor that performs an operation to estimate the author of a target image using the AI ​​model generated above.

9. In Paragraph 8, The above-mentioned backbone network is an AI model-based art style information generation and artist inference system, which is a model pre-trained to extract feature information of the shape and color of an image from a predetermined dataset.

10. In Paragraph 8, The above Gram matrix includes correlations between the extracted feature maps, and The above processor is, An AI model-based art style information generation and author inference system that performs expression learning to generate the unique style feature vector based on the correlations included in the above Gram matrix.

11. In Paragraph 10, The above processor is, An AI model-based art style information generation and artist inference system that performs the above-mentioned expression learning using a loss function to distinguish unique style feature vectors generated from paintings by different artists.

12. In Paragraph 11, The above processor is, An AI model-based art style information generation and author inference system that performs the above classification learning using a loss function for estimating the author from a unique style feature vector obtained from an expression layer that performs the above expression learning.

Images