Computational analysis of the 20th Century Korean paintings

Computational analysis of the 20th Century Korean paintings | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Computational analysis of the 20th Century Korean paintings Seohyun Baek, So-Jeong Park, So-Eun Park, You-Min Im, Jongwon Choi, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7315560/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 31 Jan, 2026 Read the published version in npj Heritage Science → Version 1 posted 21 You are reading this latest preprint version Abstract This study presents a machine learning-based approach for analyzing and classifying modern and contemporary Korean paintings using image data. The model leverages a visual-language multimodal model for efficient visual feature extraction and employs a multi-layered image analysis to capture detailed formal characteristics. Color features are extracted through analyses of various color spaces, while texture information is quantified within the texture feature space. The extracted feature vectors are analyzed and visualized through clustering, achieving an artist classification accuracy of 82.4%. Representative images from each artist cluster effectively encapsulate and highlight distinctive color and textural characteristics. Additionally, image captioning techniques were applied to generate textual descriptions of the representative images, successfully translating visual features into descriptive text. The findings confirm that machine learning-based image analysis offers an effective and objective methodology for identifying and classifying the unique characteristics of modern and contemporary Korean paintings. Artist identification Computational aesthetics Modern and contemporary paintings Painting feature extraction Unsupervised learning Figures Figure 1 Figure 2 Figure 3 1 Introduction Analyzing artworks is inherently complex 1,2 . Art experts describe visual features such as space, texture, edges, form, shape, color, composition, lighting, brushstroke, tone, and line 3,4 . They also assess movement, harmony, balance, contrast, proportion, and pattern 5 . These elements provide quantifiable data and insights into an artist’s techniques, intentions, and narratives. Each artwork carries a unique signature 6 . This distinctive visual style helps identify artistic connections and explain art movements and genres 7 . For instance, quantitative analysis of brushstroke configurations serves as a strong indicator of a painter’s style 8 . However, defining such styles can be challenging, as they often overlap, and artists may adopt multiple styles, complicating stylistic recognition 9,10 . For example, Pablo Picasso painted in both Surrealist and Cubist styles, continuously evolving over time. Moreover, human judgment in art is highly subjective and often controversial 11,12 , as experts rely heavily on personal experience and knowledge 13 . The collaboration between art and technology has a long history, with science supporting art experts since the early 18th century 14 . The development of computerized image processing techniques, including ultraviolet fluorescence, infrared reflectography, stereo microscopy, and X-radiography 15 , along with machine learning and computer vision algorithms 16 , has enabled advanced computational analysis as an interdisciplinary tool for examining paintings and deepening our understanding of art 17,18 . Additionally, the growing availability of large image datasets, such as WikiArt and ImageNet 19 , has further expanded the possibilities for computational art analysis 4,20 . Computational algorithms are valuable for identifying and comparing artistic styles and similarities between paintings. Machine learning models can encode. discriminative visual features 21 and serve as effective tools for detecting forgeries and authenticating unknown artworks 1,21,22 . Additionally, deep neural networks can uncover hidden patterns, signatures, and meaningful relationships among artworks while classifying painting styles 9,12 . Studies suggest these computational methods often surpass even highly trained art experts in accuracy 23 , reinforcing computational aesthetics as a powerful approach for analyzing visual properties in the era of digital art history 24,25 . Research on AI-based art classification has taken various approaches, primarily utilizing Convolutional Neural Networks (CNN) 26-28 and Attention Mechanisms 29-31 . Traditional machine learning methods have also been employed, combining feature extraction techniques such as Principal Component Analysis (PCA) 32 or Histogram of Oriented Gradients (HOG) 33 with K-means clustering 34-36 . However, most of these studies focus on Western artworks 37 , leading to a bias toward specific art movements, while research on traditional Korean paintings remains relatively scarce. 2 Related work In this study, we conduct an extensive literature review, identifying key trends and characteristics in the computational analysis of artworks. Additionally, we provide a comprehensive scholarly examination of representative modern and contemporary Korean painters, analyzing their distinctive styles and historical significance within the broader context of art history. Classification of art historical movements The majority of studies on computational analysis have focused on the automatic classification of artworks based on categories such as artist, style, or genre. Several studies have specifically addressed automatic artist classification and identification 1,10 , style classification 5,9 , and genre classification 17,38 . Previous studies on artist identification, artistic style, and art movement classification have extracted various features, including handcrafted features for classical machine learning 6,39 and deep features to characterize digitized paintings for deep learning 1,10,40 . Regarding handcrafted features, some researchers have utilized low-level color and texture features 2,41,42 , while others have employed high-level semantic features 43 or a combination of both 5 . In studies of Vincent van Gogh’s paintings, quantitative features such as brushstroke distribution, orientation, width, length, color palette, composition, and shape have contributed to defining a unique signature of the artist’s style and aiding artist identification 14,41,44 . Focusing on classification models Recent classification studies have primarily used Generative or Discriminative models, or both 30,45,46 . Some focus exclusively on Generative models, while others rely solely on Discriminative models. Additionally, Graphical models have been explored for representation learning, and Hypothesis Matching models have been applied for data analysis. The classification of artworks using CNNs depends on extracting meaningful features while maintaining efficiency. Recent advancements in Attention Mechanisms allow models to capture detailed stylistic elements and distinguish subtle differences between art styles, addressing limitations in conventional CNN models 30,45,46 . By selectively emphasizing relevant image regions, attention-based approaches improve classification accuracy. The extraction of color and texture features is essential for distinguishing artistic styles, as different styles exhibit unique color schemes and brushwork. Efficiently integrating these features with neural networks enhances learning and classification accuracy in automated art analysis 47-49 . The Multi-Class Kernel Method improves classification robustness, especially for figurative styles with diverse feature components such as chromatic properties, textures, morphology, and composition. By mapping these features into high-dimensional spaces, this method captures complex, non-linear relationships, enabling more precise differentiation between artistic styles 50 . Additionally, transfer learning leverages pre-trained models to enhance performance on new tasks by transferring knowledge from large-scale datasets, reducing the need for extensive computational resources 45 . Many researchers have used WikiArt 51 , one of the most comprehensive large-scale datasets, containing 150,000 artworks from 2,500 artists 8,10 . Other datasets include ArtCyclopedia 52 , Artstor Digital Library 43 , BBC Painting Dataset 53 , Mark Harden’s Artchive 5 , ABC Gallery 9 , and Artlex & CARLI Digital Collections 7 . Some studies also gathered images from Wikipedia, Flickr, online museums, and internet sources 54,55 . Park et al. 56 selected images from 25 artists in the WikiArt dataset and addressed class imbalance using a weighted cross-entropy loss function. They expanded the dataset by applying augmentation techniques (e.g., resizing, flipping, rotation via Albumentations). CLAHE improved contrast, while CutMix enhanced texture information. By fine-tuning ResNet50, adjusting fully connected layers, and freezing convolutional weights, they significantly improved artist classification accuracy. Previous studies have explored the classification of traditional Chinese paintings (TCP) 57-59 . Li & Wang 57 used wavelets and 2D Multi-resolution Hidden Markov Models (MHMM) to categorize Chinese ink paintings by style and artist. Jiang et al. 58 distinguished TCP from non-TCP images and classified them into Gongbi (skilled brush, 1,889 images) and Xieyi styles using low-level features (color, texture, edge) and a hybrid classifier (decision tree + SVM). Their approach achieved practically viable accuracy. Lu et al. 59 developed a TCP classification framework based on four art movements (Xieyi, Gongbi, Goule, Shese) and six painters. They applied Bayesian classifiers, k-NN, fuzzy C-means clustering, and a non-linear multi-class SVM, comparing their classification performance. Recent studies have increasingly focused on improving artwork classification accuracy through data augmentation and advanced deep learning models. Baldrati et al. 60 introduced a CLIP-based multi-modal approach using the NoisyArt dataset, combining textual and visual features to enhance classification and retrieval tasks. This study demonstrated the effectiveness of multi-modal learning in computational art analysis. Zhong et al. 61 proposed a Two-Channel Dual-Path Network (FPTD) that incorporates RGB and brushstroke texture information to improve fine-art painting classification. The study used a Gray-Level Co-Occurrence Matrix (GLCM) to extract texture details from multiple directions, enabling more precise classification of styles, artists, and genres while enhancing generalization performance. Kim et al. 62 developed a proxy learning method that integrates pre-trained language models with visual data for analyzing artistic styles. By modeling the relationships between textual descriptions and visual features, their method extracts meaningful visual concepts, significantly improving automated classification and analysis of artworks. This interdisciplinary approach broadens traditional feature extraction techniques, offering new insights into computational art analysis. Modern and contemporary Korean artists Kim Ki-chang, Kim Whan-ki, Do Sang-bong, Park Soo-keun, Yoo Young-kuk, Lee Jung-seob, Chun Kyung-ja, Chang Uc-chin, Byun Kwan-sik, Lee Sang-beom, and Byun Jong-ha are key figures in modern and contemporary Korean art. Active in the 20th-century Korean art scene, they played a crucial role in shaping Korean modernism and establishing the national artistic identity in the global art world. Amidst the turbulent history of modern Korea, these artists developed distinctive artistic styles, producing works that reflect the evolution of Korean art. Among them, Kim Whan-ki, Park Soo-keun, Yoo Young-kuk, Lee Jung-seob, and Chang Uc-chin are regarded as the five leading second-generation Western-style painters and first-generation modernists in Korean art 63 . Kim Whan-ki (1913–1974) and Yoo Young-kuk (1916–2002) explored Korean modernity through experimental abstractions that bridged tradition and contemporary life. While Yoo Young-kuk pursued pure geometric abstraction, continuing pre-war abstract movements 64 , Kim Whan-ki integrated real-world motifs to express the spirit of the times 65 . Lee Jung-seob (1916–1956), known for his restrained color palette, depicted local sentiments through motifs such as cows and children 66 . Chang Uc-chin (1917–1990) explored formative art inspired by his hometown imagery. Park Soo-keun (1914–1965) captured ordinary people’s daily lives during the Japanese colonial period and the Korean War, expressing their resilience and humanity with a “sincere heart and kind gaze” 67 . Except for Park Soo-keun, the other four artists were members of the New Realism Group (Shinsasilpa), an influential painters' collective founded in July 1947 68 . Comprising mostly graduates of Japanese art schools, the group sought to merge modernist principles with traditional Korean identity, responding to the post-liberation and wartime political upheavals. Kim Ki-chang (1914–2001) was an Oriental painter known for his "Foolish Painting Style," which modernized traditional Joseon Dynasty folk and genre paintings 69 . Do Sang-bong (1902–1977) integrated Korean sentiment into traditional realist painting from the late 1920s to the 1970s, employing balanced compositions and rich color palettes 63 . Chun Kyung-ja (1924–2015), the only female artist among the eleven, forged a distinctive style by blending traditional Korean colors with bold, diverse hues. In an era dominated by ink wash painting, she introduced new possibilities for colored painting, merging Korean sentiment with modern aesthetics 70 . Byun Kwan-sik (1899–1976), active from the 1920s to the mid-1970s, upheld the tradition of Korean painting while pioneering the "Sojeong style," which featured distinctive compositions and varied ink techniques 71 . Lee Sang-beom (1897–1972), influenced by photography and Western painting from 1923 onward, became a master of ink wash landscape painting. He developed the "Cheongjeon style," shifting from conceptual landscapes to depictions of real scenery, incorporating mijeomjun techniques to modernize traditional landscape painting 72 . Byun Jong-ha (1926–2000), active from the 1940s to 2000, developed a "three-dimensional painting" style in the 1960s, introducing depth and spatiality to flat paintings. By blurring the boundary between painting and sculpture, his work became a source of inspiration 73 . 3 Methodology Image datasets In this study, we constructed a dataset comprising 1,100 images from 11 representative modern and contemporary Korean painters, with 100 images per artist (Table 1). These were sourced primarily from the MMCA, which, as of October 28, 2024, holds 11,479 images, including 3,608 modern paintings. To expand the dataset, we gathered additional high-resolution images from reputable sources such as Google Arts & Culture and Google Search. We ensured authenticity by selecting images from authoritative websites that provided verified artwork details, including the title and year of creation. Using Google Search, we applied the 'artist's name' as a keyword and the 'large size' filter to obtain high-resolution images. For missing metadata (e.g., title, creation year), we utilized Google Lens to perform reverse image searches, retrieving verified details from artist foundation websites and other credible sources. When low-resolution images were the dataset with verified artworks, ensuring data reliability and comprehensive coverage of modern and contemporary Korean paintings. Framework This study proposes an analytical framework that integrates multiple image feature extraction techniques, dimensionality reduction, and clustering to analyze artworks by modern and contemporary Korean painters (Figure 1). The framework consists of the following steps: First, input data comprises images categorized by artist, where object regions within images are cropped based on pre-existing annotation data before processing. The collected artwork images are then transformed into diverse visual feature vectors using multiple image encoders. Specifically, four encoding methods—RGB means 74 , HSV means 48 , color histograms, Gray-Level Co-occurrence Matrix (GLCM) 75 for texture analysis, and CLIP embeddings 76 —are employed to extract distinct feature representations. Among these, RGB and HSV were employed as color spaces, while GLCM served as the texture feature space. These features are subsequently normalized and concatenated to construct a comprehensive multimodal feature vector for each image. Each feature vector is stored along with its corresponding filename for reference. The feature vectors of all images are then structured into a feature matrix, encapsulating the visual characteristics of the entire dataset. To facilitate visualization of high-dimensional features in a low-dimensional space, we apply t-SNE, while K-means clustering is used to identify typological similarities and group characteristics among the artwork images. Upon image input, feature vectors are generated through four modules: RGB, HSV, GLCM, and CLIP. The RGB module extracts color composition via red, green, and blue channels, while the HSV module computes mean and standard deviation values for each color component based on edge count, dark pixels, symmetry, and average values in hue, saturation, and value spaces. GLCM, a statistical texture analysis method, characterizes textures by quantifying the frequency of specific pixel pair occurrences at predefined spatial relationships, enabling detailed statistical measurements. Lastly, CLIP learns semantic associations between text and images, encoding input images into embedding vectors and linking them with textual descriptions. The model calculates cosine similarity between image and text embeddings, selecting the most relevant caption. This study employs pre-trained CLIP encoders for improved feature representation. The clustering module employs K-means clustering on low-dimensional feature vectors projected by t-SNE, automatically grouping modern and contemporary Korean paintings by type. Feature vectors from each module are concatenated into a unified vector per image, which is then vertically stacked across all images to form a feature matrix. This high-dimensional matrix is reduced to two dimensions—x and y axes—via t-SNE, enabling visual representation. By integrating the four modules, the matrix effectively captures key color and texture attributes, aiding in classification. The two-dimensional vectors, preserving essential visual features in a lower-dimensional space, serve as input for clustering. Subsequently, K-means clustering partitions the dataset into 11×20 initial clusters. The K value was chosen to account for the diverse styles and evolving techniques of the 11 modern and contemporary Korean painters in the dataset. These initial clusters aim to capture fine stylistic variations, maximizing granularity in the early stages. Post-processing was applied to refine clustering by analyzing cluster sizes, removing smaller clusters, and retaining only the upper half to enhance interpretability. Final clustering was then performed by using the central coordinates of the retained clusters and reapplying K-means. The clustering results were evaluated by selecting representative images for each cluster and analyzing the distribution of ground truth labels (painter names) to assess cluster purity and average accuracy. The most frequently occurring ground truth label within each cluster was assigned as its representative label, quantifying how well clusters correspond to individual painters' stylistic tendencies. Classification outcomes were visualized using representative images and captions, providing interpretable insights into model performance. Additionally, clusters were displayed with circle colors indicating ground truth labels and outline colors representing clustering results, enabling intuitive verification of label alignment. This clustering strategy balances capturing artistic diversity in the initial stages while refining clusters for interpretability. Given that modern and contemporary Korean painters exhibit both distinctive personal styles and intra-artist variations across periods and subjects, this hierarchical clustering methodology effectively accommodates the dataset’s characteristics. 4 Experiments Implementation detail In this study, we constructed feature vectors by extracting image embeddings using the pre-trained CLIP model ViT-B/32 and integrating additional color and texture information. The experiments were conducted on an NVIDIA RTX 3090 GPU with Python 3.10, running on Ubuntu 22.04 with CUDA 12.2. Input images were normalized to 256×256 resolution, and a batch size of 32 was set for CLIP embedding extraction. Feature vector construction included one-dimensional RGB and HSV means per channel, 768-dimensional color histograms (256 per channel), and 23-dimensional GLCM features based on co-occurrence matrices with 23 gray levels and a 1-pixel offset. Additionally, 529-dimensional LBP features were generated using 23×23 patches with a radius of 1, 8 neighbors, and bilinear resampling. These features were concatenated sequentially with CLIP embeddings, forming a final 1338-dimensional feature vector, with z-score normalization applied before clustering. For t-SNE-based dimensionality reduction, a learning rate of 200 and perplexity of 30 were set. K-Means clustering was initially performed with 20× the number of artists, followed by filtering to retain the top 50% of clusters by size. Clustering accuracy was assessed based on the proportion of majority labels within each cluster. To ensure experimental reproducibility, random seeds were fixed for all t-SNE and K-Means processes. Result and discussion We summarize the clustering results in Table 2, with the best results highlighted in bold. Data clustering was performed based on the visual characteristics of images, and the consistency within each cluster was assessed through intuitive visualization and quantitative evaluation. Specifically, visual features extracted from the pre-trained CLIP model were combined with RGB, HSV, histogram-based color features, and texture features derived from LBP (Local Binary Pattern) and GLCM. These features were projected into a lower-dimensional space using t-SNE and clustered into 11 primary groups using the K-Means algorithm. Table 2. Classification accuracy results by method for artist identification (unit: %). Dataset Method RGB HSV Histogram LBP GLCM Ours Korean modern painting 82.0 81.3 51.0 68.8 73.7 82.4 Quantitative performance evaluation was conducted by assigning the most frequently occurring label within each cluster as the representative label and defining accuracy as the proportion of images matching this label. This metric effectively quantifies cluster purity and is particularly suitable when ground truth labels are clearly defined (e.g., classification by artist or artwork type). Additionally, it enabled direct comparison between our method and previous feature extraction techniques. Prior studies primarily relied on RGB, HSV, and histogram-based features 77 for color extraction, while texture analysis was performed using LBP and GLCM to capture stylistic elements such as brushstroke patterns. Our method, which integrates these traditional features with a comprehensive visual-semantic representation, outperformed previous approaches, achieving an accuracy of 82.4% under the same dataset and experimental conditions. This result suggests that our approach more effectively captures the visual characteristics of artworks by complementing conventional color and texture features with richer semantic information. While quantitative performance evaluation is standard in traditional machine learning research, qualitative interpretation often dominates studies in artistic data and visual cultural heritage. However, this study incorporates essential quantitative clustering accuracy measurements to ensure the objectivity of machine learning-based analysis and enhance comparability with future research. This approach goes beyond merely reporting classification accuracy; it serves as a validation process to determine whether each cluster is meaningfully structured around distinct visual characteristics. Through additional analysis, we compared the performance of traditional color-and texture-based methods with our proposed approach, categorized by artist and artwork type, as summarized in Table 3. Color-based models, such as RGB and HSV, exhibited high accuracy for artists known for their distinctive color usage. For example, Yoo Young-kuk, who frequently employed bold primary colors and geometric color planes, achieved 85.0% accuracy with RGB and 92.0% with HSV. Similarly, Kim Whan-ki recorded 72.0% and 75.0% accuracy in RGB and HSV, respectively. Do Sang-bong also demonstrated exceptionally high recognition rates in color-based analysis due to his consistent color schemes in still life paintings, characterized by recurring tonalities and specific color harmonies. These results suggest that color-based analytical methods are particularly effective when an artist’s signature color palette is a defining feature of their work. Table 3. Comparison of model accuracy by artist (unit: %). Dataset Method RGB HSV Histogram LBP GLCM Ours Kim Ki Chang 41.0 44.0 33.0 34.0 42.0 73.5 Kim Whan Ki 72.0 75.0 41.0 66.0 76.0 74.8 To Sang Bong 98.0 98.0 37.0 98.0 100 97.0 Park Soo Keun 56.0 56.0 38.0 50.0 56.0 69.1 Byeon Gwan Sik 50.0 49.0 30.0 53.0 51.0 76.0 Byun Jong Ha 49.0 58.0 24.0 36.0 51.0 78.3 Yoo Young Kuk 85.0 92.0 56.0 70.0 93.0 94.9 Lee Sang Beom 65.0 64.0 32.0 63.0 60.0 78.4 Lee Jung Seop 43.0 42.0 27.0 26.0 40.0 74.4 Chang Uc Chin 49.0 47.0 24.0 36.0 60.0 86.4 Cheon Gyeong Ja 67.7 75.8 18.2 65.7 74.8 82.3 Conversely, histogram-based methods showed relatively low accuracy across all cases, indicating that simple color distribution analysis alone is insufficient to capture the nuanced stylistic and painterly characteristics of each artist. Notably, Byun Jong-ha and Chang Uc-chin recorded accuracy rates below 30%, highlighting the limitations of relying solely on basic color frequency comparisons for artist identification. The LBP method, which learns patterns by analyzing image textures and tonal variations, demonstrated high performance for Do Sang-bong, likely due to the detailed brushwork and shading characteristic of his still life paintings. Yoo Young-kuk also showed strong performance, benefiting from the high contrast between color fields in his works. Meanwhile, the GLCM method, which captures patterns by analyzing spatial relationships between pixels, achieved 100% accuracy for Do Sang-bong, attributed to the consistent composition and recurring patterns in his still life paintings. In contrast, artists such as Kim Ki-chang and Lee Jung-seob, whose works feature more spontaneous and dynamic compositions, exhibited relatively lower performance in pattern-based analysis. Similarly, the experimental and complex techniques of Byun Jong-ha, as well as the simplified and flat compositions characteristic of Chang Uc-chin, were not optimally captured by any specific method, leading to overall lower accuracy rates for these artists. These findings suggest that color-based techniques are more effective for artists who emphasize distinctive color usage, while texture and pattern analysis methods perform better for artists whose works focus on compositional structure and surface textures. This underscores the need for selecting analytical methods based on the inherent visual characteristics of each artist’s body of work. In summary, conventional methods demonstrated relatively high accuracy for certain artists but showed significant performance degradation for those employing distinctive or unconventional expressive techniques. This limitation stems from traditional methods' sensitivity to specific color combinations or surface texture patterns, which prevents them from fully capturing the holistic visual context of artworks. In contrast, the proposed method, which integrates CLIP-based features with color and texture information, achieved stable and consistent accuracy across all artists and artwork categories. By comprehensively reflecting diverse visual characteristics without bias toward specific artists or styles, this approach overcomes the limitations of conventional methods. This result is visualized in Figure 2. Using K-means clustering, the images were grouped into 11 main clusters, distinguished by color variations. Each artwork is represented as a point, with dark-bordered points indicating correctly classified images and lighter-bordered points representing misclassifications, allowing for easy verification of accuracy. The high proportion of dark-bordered points suggests that, despite being an unsupervised learning model, the system effectively identifies similarities among artworks, accurately classifying them by artist. This highlights the model’s strong ability to distinguish artists based solely on image information, underscoring the value of AI-driven analysis. Figure 3 presents a schematic visualization of representative images closest to each cluster centroid. These images encapsulate the dominant visual characteristics of their respective clusters, which are grouped based on stylistic attributes such as color composition, texture patterns, and compositional structures. For instance, one cluster primarily consists of traditional ink paintings with monochromatic tonal variations and brushwork textures, while another is dominated by abstract works featuring bold primary colors and geometric forms. Still life paintings form a distinct cluster as well. This visualization confirms that the clustering process successfully captures the latent visual attributes of each artwork. By inspecting representative images within each cluster, it is evident that the proposed analytical approach effectively groups visually similar works, offering an interpretable mapping of stylistic relationships between artists and artworks. Additionally, to analyze differences between groups, the generated cluster centroids were examined. These centroids represent the average position of data points within each cluster, effectively summarizing the dominant color tones and texture characteristics of each group. Table 4 presents the centroids of correctly classified clusters, demonstrating that the generated captions accurately reflect the visual attributes of each group. This suggests that both the clustering and caption generation models exhibit high reliability in capturing distinctive visual features. However, as shown in Table 5, certain clusters produced incorrect classifications and captions, highlighting areas for further refinement. 5 Conclusion This study introduces a novel analytical framework for analyzing and classifying artworks by modern and contemporary Korean painters, integrating color, texture, and semantic features. While previous research has largely focused on Western paintings, studies on Eastern art—particularly Korean paintings—remain limited. To address this gap, we constructed a dedicated dataset of works by major Korean artists and developed a methodology that captures their unique visual language and aesthetic characteristics. Comparative analysis of various feature extraction methods reveals that artists with distinctive color palettes achieve high accuracy in color-based analysis, while those emphasizing texture and brushwork perform better in texture-based approaches. In contrast, histogram-based methods relying solely on color frequency distributions struggle to capture unique artistic expressions. To overcome these limitations, this study presents an integrated approach that combines CLIP-based semantic features with visual attributes such as color and texture. The proposed method ensures stable and balanced performance across different artists and styles, effectively identifying visual similarities even in an unsupervised learning environment. The clustering results and representative image analysis confirm that the proposed framework effectively distinguishes artworks by artist. Representative images summarize each artist’s visual language in terms of color tone, texture, and composition, demonstrating that the framework not only enhances classification accuracy but also serves as a quantitative tool for identifying visual characteristics, exploring inter-artwork relationships, and situating works within broader art historical contexts. Additionally, this study contributes to the digital analysis of Korean modern and contemporary paintings by laying the groundwork for future artist classification models and stylistic analysis methods tailored to Korean art. It addresses the limitations of prior research, which predominantly focused on Western art while overlooking the unique visual culture and historical context of Korean paintings. In conclusion, this study introduces a novel AI-driven methodology that integrates visual analysis with semantic interpretation for the study of Korean modern and contemporary paintings. The experimental results highlight its effectiveness and academic significance, quantitatively capturing the distinct visual language and formal characteristics of Korean paintings. The dataset and analytical framework developed here offer a foundation for future research in digital art history. Expanding the dataset to include traditional Korean paintings and postmodern works would enable a more comprehensive chronological analysis, further illuminating the stylistic evolution and defining characteristics of Korean art across different periods and artists. Declarations Availability of data The dataset used in this study consists of paintings by eleven modern and contemporary Korean artists. The images were provided by various institutions, including the National Museum of Modern and Contemporary Art (MMCA), the Jang Uc-Chin Foundation, Seoul Metropolitan Government, the Park Soo Keun Institute, and the Yoo Youngkuk Art Foundation. The copyright for the images belongs to the respective institutions. Competing interests The authors declare that they have no competing interests. Acknowledgements This research was supported by Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2023 (Project Name: Acquisition of 3D precise information of microstructure and development of authoring technology for ultra-high precision cultural restoration, Project Number: RS-2023-00227749, Contribution Rate: 100%) Authors' contributions SH took primary responsibility for drafting the manuscript, designing the experiments, and conducting the main model experiments. SJ, SE, and YM contributed to data collection, data refinement, experimental interpretation, and the review of prior research. JW was responsible for the experimental design and interpretation of the results. BA supported the data refinement, validation of experimental results, and overall quality assurance. All authors read and approved the final manuscript. Authors' information Seohyun Baek: First Author, Electronics and Telecommunications Research Institute (ETRI), Daejeon, Korea. E-mail: [email protected] So-Jeong Park: Contributing Author, Chung-Ang University, Seoul, Korea. E-mail: [email protected] So-Eun Park: Contributing Author, Chung-Ang University, Seoul, Korea. E-mail: [email protected] You-Min Im: Contributing Author, Chung-Ang University, Seoul, Korea. E-mail: [email protected] Jongwon Choi: Contributing Author, Chung-Ang University, Seoul, Korea. E-mail: [email protected] Bo-A Rhee: Corresponding Author, Chung-Ang University, Seoul, Korea. E-mail: [email protected] References Viswanathan, N. & Stanford. Artist identification with convolutional neural networks. https://api.semanticscholar.org/CorpusID:198974841 (2017). Kim, M. & Kim, J. Complementary quantitative approach to unsolved issues in art history: Similarity of visual features in the paintings of Vermeer and his probable mentors. Leonardo 52 , 164-174 (2019). Shamir, L. et al. Impressionism, expressionism, surrealism: Automated recognition of painters and schools of art. ACM Transactions on Applied Perception 7 , 1-17 (2010). Cetinic, E., Lipić, T. & Grgic, S. Fine-tuning convolutional neural networks for fine art classification. Expert Syst. Appl 114, 107-118 (2018). Saleh, B., Abe, K., Arora, R. & Elgammal, A. Toward automated discovery of artistic influence. Multimedia Tools and Applications 75 , 3565-3591 (2016). Bar, Y., Levy, N. & Wolf, L. Classification of artistic styles using binarized features derived from a deep neural network. In Proc. ECCV Workshops , 71-84 (2014). Zujovic, J. et al. Classifying paintings by artistic genre: An analysis of features & classifiers. In Proc. IEEE international workshop on multimedia signal processing (2009). Saleh, B. & Elgammal, A. A unified framework for painting classification. In Proc. IEEE International Conference on Data Mining Workshop (2015). Gultepe, E., Conturo, T. E. & Makrehchi, M. Predicting and grouping digitized paintings by style using unsupervised feature learning. Journal of cultural heritage 31 , 13-23 (2018). Chen, J. & Deng, A. Comparison of machine learning techniques for artist identification. Standford193CS231N Report ; Art Computer Science: Stanford, CA, USA (2018). Berezhnoy, I. E., Postma, E. O. & van den Herik, H. J. Authentic: computerized brushstroke analysis. In Proc. IEEE International Conference on Multimedia and Expo (2005). van den Herik, H. J. & Postma, E. O. Discovering the visual signature of painters. In Future Directions for Intelligent Systems and Information Sciences: The Future of Speech and Image Technologies. Brain Computers, WWW, and Bioinformatics 129-147 (2000). Falomir, Z., Lledó, M., Sanz, I. & Gonzalez-Abril, L. Categorizing paintings in art styles based on qualitative color descriptors, quantitative global features and machine learning (QArt-Learn). Expert Systems with Applications 97 , 83-94 (2018). Lamberti, F., Sanna, A. & Paravati, G. Computer-assisted analysis of painting brushstrokes: digital image processing for unsupervised extraction of visible features from van Gogh’s works. EURASIP Journal on Image and Video Processing 1-17 (2014). Johnson, R. C. et al. Image processing for artist identification. IEEE Signal Processing Magazine 25 , 37-48 (2008). van der Maaten, L. & Postma, E. Identifying the real van gogh with brushstroke textons. White Paper , Tilburg University (2009). Condorovici, R. G., Vertan, C. & Florea, L. Artistic genre classification for digitized painting collections. UPB Scientific Bulletin 75 , 75-86 (2013). Stork, D., Coddington, J. & Bentkowska-Kafel, A. Computer vision and image analysis of art. In Proceedings of SPIE 7531 (2010). Deng, J. et al. Imagenet: A large-scale hierarchical image database. In Proc. IEEE conference on computer vision and pattern recognition (2009). Sigaki, H. Y. D., Perc, M. & Haroldo, R. V. History of art paintings through the lens of entropy and complexity. National Academy of Sciences 115 , E8585–E8594 (2018). Elgammal, A. et al. The shape of art history in the eyes of the machine. In Proc. AAAI Conference on Artificial Intelligence 32 , 2183-2191 (2018). Nemade, R., Nitsure, A., Hirve, P. & Mane, S. B. Detection of forgery in art paintings using machine learning. Int. J. Innov. Res. Sci. Eng. Technol 6 , 8681-8692 (2017). Stork, D. Computer image analysis of paintings and drawings: an introduction to the literature. In Proc. Image Processing for Artist Identification Workshop (2008). Cetinic, E., Tomislav, L. & Grgic, S. A deep learning perspective on beauty, sentiment, and remembrance of art. IEEE Access 7 , 73694-73710 (2019). Ellis, M. H. & Johnson Jr., C. R. Computational connoisseurship: enhanced examination using automated image analysis. Visual Resources 35 , 125-140 (2019). Yu, Q. & Shi, C. An image classification approach for painting using improved convolutional neural algorithm. Soft Computing 28, 847-873 (2024). Mondal, K. & Anita, H. Categorization of artwork images based on painters using CNN. J. Phys.: Conf. Ser. (2021). Cervantes, F. et al. Evaluation of CNN models with transfer learning in art media classification in terms of accuracy and class relationship. Computación y Sistemas 28, 233-244 (2024). Zhang, X. & Ding, T. Style classification of media painting images by integrating ResNet and attention mechanism. Heliyon 10 , e27178 (2024). Li, Q., Yue, L. & Feng, Q. Research on automatic classification method of artistic styles based on attention mechanism convolutional neural network. Int. J. Low-Carbon Technol. 19 , 2018-2023 (2024). Zeng, L. Improved painting image style classification of ResNet based on attention mechanism. In Proc. International Conference on Advances in Electrical Engineering and Computer Applications (2023). Polak, A. et al. Hyperspectral imaging combined with data classification techniques as an aid for artwork authentication. J. cultural heritage 26 , 1-11 (2017). Sahu, T., Tyagi, A., Kumar, S. & Mittal, A. Classification and aesthetic evaluation of paintings and artworks. In Proc. 13th International Conference on Signal-Image Technology & Internet-Based Systems (2017). Castellano, G. & Vessio, G. A deep learning approach to clustering visual arts. Int. J. Comput. Vis. 130 , 2590-2605 (2022). Abas, F. S. & Martinez, K. Classification of painting cracks for content-based analysis. In Machine Vision Applications in Industrial Inspection XI 5011 , 149-160 (SPIE, 2003). Andronache, I. Can complexity measures with hierarchical cluster analysis identify overpainted artwork? Zenodo (2023). Baek, S. et al. Identification of Korean Neo-Realism artists through CLIP-based analysis. J. Korea Soc. Comput. Inf. 29 , 317-328 (2024). Agarwal, S., Karnick, H., Pant, N. & Patel, U. Genre and style based painting classification. In Proc. IEEE Winter Conference on Applications of Computer Vision (2015). Mensink, T. & Van Gemert, J. The rijksmuseum challenge: Museum-centered visual recognition. In Proc. International conference on multimedia retrieval (2014). Kim, D., Xu, J., Elgammal, A. & Mazzone, M. Computational analysis of content in fine art paintings. In Proc. ICCC (2019). Li, J., Yao, L., Hendriks, E. & Wang, J. Z. Rhythmic brushstrokes distinguish van Gogh from his contemporaries: findings via automated brushstroke extraction. IEEE Trans. Pattern Anal. Mach. Intell. 34 , 1159-1176 (2011). Park, J. W. Secrets of balanced composition as seen through a painter's window: visual analyses of paintings based on subset barycenter patterns. In ACM SIGGRAPH 2019 Art Gallery 1-10 (2019). Carneiro, G. Graph-based methods for the automatic annotation and retrieval of art prints. In Proc. 1st ACM International conference on multimedia retrieval (2011). Dasgupta, S. & Hossain, M. A. Pattern recognition through brushstroke analysis: an inspection to classify van Gogh from others. BRAC University , (2017). Xu, K. Automated artist identification using deep learning and transfer learning techniques. Sci. Technol. Eng. 1 , 8 (2024). Zheng, S. Deep learning-based artistic style transformation algorithm in visual communication. Int. J. Syst. Assurance Eng. Manag. 1-10 (2024). Wei, N. Research on the algorithm of painting image style feature extraction based on intelligent vision. Futur. Gener. Comput. Syst. 123 , 196-200 (2021). Nunez-Garcia, I. et al. Classification of paintings by artistic style using color and texture features. Comput. Sist. 26 , 1503-1514 (2022). Hu, B. & Yang, Y. Construction of a painting image classification model based on AI stroke feature extraction. J. Intell. Syst. 33 , 20240042 (2024). Wang, Q. Painting style recognition and classification based on convolutional neural network and multi-kernel learning. In Intelligent Computing Technology and Automation 1059-1066 (2024). WikiArt. WikiArt Visual Art Encyclopedia . n.d. Web. 3 July 2025. http://www.wikiart.org. Ivanova, K. et al. Features for art painting classification based on vector quantization of mpeg-7 descriptors. In Proc. Data Engineering and Management: Second International Conference 178-189 (2012). Crowley, E. J. & Zisserman, A. In Search of Art. In Agapito, L., Bronstein, M. M. & Rother, C. (eds.) Computer Vision - ECCV 2014 Workshops: Zurich, Switzerland, September 6-7 and 12, 2014, Proceedings, Part I 54-70 (2015). Shamir, L. Computer analysis reveals similarities between the artistic styles of Van Gogh and Pollock. Leonardo 45 , 149-154 (2012). Karayev, S. et al. Recognizing image style. arXiv preprint arXiv:1311.3715 (2013). Park, J.S., Kim, S.Y., Yoon, Y.C. & Kim, S. K. Optimizing CNN structure to improve accuracy of artwork artist classification. J. Korea Soc. Comput. Inf. 28 , 9-15 (2023). Li, J. & Wang, J. Z. Studying digital imagery of ancient paintings by mixtures of stochastic models. IEEE Trans. Image Process. 13 , 340–353 (2004). Jiang, S., Huang, Q., Ye, Q. & Gao, W. An effective method to detect and categorize digitized traditional Chinese paintings. Pattern Recognit. Lett. 27 , 734–746 (2006). Lu, G. et al. Content-based identifying and classifying traditional chinese painting images. In Proc. Congress on Image and Signal Processing (2008). Baldrati, A. et al. Exploiting CLIP-based multi-modal approach for artwork classification and retrieval. arXiv preprint arXiv:2309.12110 (2023). Zhong, S., Huang, X. & Xiao, Z. Fine-art painting classification via two-channel dual path networks. Int. J. Mach. Learn. Cybern. 11 , 137–152 (2020). Kim, D., Elgammal, A. & Mazzone, M. Formal analysis of art: proxy learning of visual concepts from style through language models. arXiv preprint arXiv:2201.01819 (2022). National Museum of Modern and Contemporary Art, Korea. The Most Honest Confession: Chang Ucchin Retrospective . n.d. Web. 14 Sept. 2023. https://www.mmca.go.kr/exhibitions/exhibitionsDetail.do?menuId=1030000000&exhId=202302150001627. Joo, S. A Study on the Sublime in YOO YOUNGKUK's Abstract Paintings: Focusing on works since the 1960s. Department of Aesthetics & Art History, Graduate School of Chosun University, Gwangju, Korea (2023). Kim, H.S. Study on the Colors of Kim Whan-ki's Painting. J. Art Theory Pract. 3 , 155–172 (2005). Kim, M.J. About Originality on Materials and Techniques in Lee Jung Seob’s Painting. J. Korean Mod. Contemp. Art Hist. 32 , 83–117 (2016). Eom, S.M. A Unique Artistic World: Park Soo-keun’s Concave and Convex Technique. One Art World 30–37 (2022). Park, J. M. The concept of abstraction in New Realism group: New interpretation on geometric abstraction from its own point of view. Korean Bulletin of Art History 38 , 240–276 (2012). Choi, B. The changing characteristics and meaning of Woonbo Kim, Ki-Chang’s ‘The Foolish Painting Style’. Oriental Art 33 , 233–256 (2016). Lee, M.S. A Study on the Cheon, Kyung-Ja's art works. Graduate School of Jeju University, Jeju, Korea . (2011). Lee, Y. U. A study on Byeon kwan sik’s Mt. Geumgang paintings. Art Hist. Cult. Herit. 1 , 9–34 (2012). National Museum of Modern and Contemporary Art, Korea. “Today, This Work: Lee Sangbeom | Early Winter | 1926.” n.d. Web. 1 June 2023. https://www.mmca.go.kr/digitals/digitalMovInfo.do?mbId=202306010001268. Hur, N.Y. The Sensibility Narrative through figuration by Byun Jong-Ha. Assoc. Biogr. Art Hist. Korea 5 , 265–287 (2009). Maxwell, J. C. Experiments on colour as perceived by the eye, with remarks on colour-blindness. Royal Society of Edinburgh 3 , 299–301 (1857). Siqueira, D., Roberti, F., Schwartz, W. R. & Pedrini, H. Multi-scale gray level co-occurrence matrices for texture description. Neurocomputing 120 , 336–345 (2013). Nukrai, D., Mokady, R. & Globerson, A. Text-only training for image captioning using noise-injected clip. arXiv preprint arXiv:2211.00575 (2022). Özlü, Ahmet. Color Recognition . GitHub, n.d. Web. 3 July 2024. https://github.com/ahmetozlu/color_recognition. Tables Tables 1 and 4 are available in the Supplementary Files section. Additional Declarations No competing interests reported. Supplementary Files Tables.doc Cite Share Download PDF Status: Published Journal Publication published 31 Jan, 2026 Read the published version in npj Heritage Science → Version 1 posted Editorial decision: Revision requested 14 Sep, 2025 Reviews received at journal 01 Sep, 2025 Reviews received at journal 01 Sep, 2025 Reviews received at journal 29 Aug, 2025 Reviews received at journal 27 Aug, 2025 Reviews received at journal 24 Aug, 2025 Reviews received at journal 17 Aug, 2025 Reviewers agreed at journal 17 Aug, 2025 Reviews received at journal 16 Aug, 2025 Reviewers agreed at journal 15 Aug, 2025 Reviewers agreed at journal 14 Aug, 2025 Reviewers agreed at journal 13 Aug, 2025 Reviewers agreed at journal 13 Aug, 2025 Reviewers agreed at journal 11 Aug, 2025 Reviewers agreed at journal 11 Aug, 2025 Reviewers agreed at journal 11 Aug, 2025 Reviewers agreed at journal 10 Aug, 2025 Reviewers invited by journal 10 Aug, 2025 Editor assigned by journal 08 Aug, 2025 Submission checks completed at journal 08 Aug, 2025 First submitted to journal 07 Aug, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7315560","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":501467939,"identity":"6c948ec3-25be-4ccd-8536-7cd114199dbd","order_by":0,"name":"Seohyun Baek","email":"","orcid":"","institution":"Electronics and Telecommunications Research Institute","correspondingAuthor":false,"prefix":"","firstName":"Seohyun","middleName":"","lastName":"Baek","suffix":""},{"id":501467941,"identity":"e0e8fb13-c90a-436f-9bfa-2179a004e820","order_by":1,"name":"So-Jeong Park","email":"","orcid":"","institution":"Chung-Ang University","correspondingAuthor":false,"prefix":"","firstName":"So-Jeong","middleName":"","lastName":"Park","suffix":""},{"id":501467942,"identity":"a3e392bf-d002-45f2-8854-205ba879f928","order_by":2,"name":"So-Eun Park","email":"","orcid":"","institution":"Chung-Ang University","correspondingAuthor":false,"prefix":"","firstName":"So-Eun","middleName":"","lastName":"Park","suffix":""},{"id":501467943,"identity":"cf1f543c-d753-4b9d-b03c-a2718297b8c4","order_by":3,"name":"You-Min Im","email":"","orcid":"","institution":"Chung-Ang University","correspondingAuthor":false,"prefix":"","firstName":"You-Min","middleName":"","lastName":"Im","suffix":""},{"id":501467944,"identity":"a9d71e94-859a-43f7-acbb-703163ee0e94","order_by":4,"name":"Jongwon Choi","email":"","orcid":"","institution":"Chung-Ang University","correspondingAuthor":false,"prefix":"","firstName":"Jongwon","middleName":"","lastName":"Choi","suffix":""},{"id":501467945,"identity":"956bf280-ede1-4da7-b4f8-aa150f33fff1","order_by":5,"name":"Bo-A Rhee","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAzElEQVRIiWNgGAWjYBACAxDB2MDAwA9iJBSQokWyAaTFgBQtBgfgXALAnL35AXPlDhs54/OrEz88MGCQ5xc7gF+LZc8xA8azZ9KMzW683SwBdJjhzNkJBBx2I8GAsbHtcOK2G2c3gLQkGNwmqCX9A1DL//rNM85u/kGklhyQLQcSDPh7txFni2XPmYKDjW3JhjNu8G6zSDCQIOwXc/b2jQ8b2+zk+fvPbr75o8JGnl+agBYQOAAmJcAqJQgrRwD+A6SoHgWjYBSMgpEEAN8QRudoK9vhAAAAAElFTkSuQmCC","orcid":"","institution":"Chung-Ang University","correspondingAuthor":true,"prefix":"","firstName":"Bo-A","middleName":"","lastName":"Rhee","suffix":""}],"badges":[],"createdAt":"2025-08-07 06:53:51","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7315560/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7315560/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s40494-026-02304-1","type":"published","date":"2026-01-31T15:58:45+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":89273631,"identity":"dcb210a6-e2c9-46fa-a2be-e32e2f6c92f3","added_by":"auto","created_at":"2025-08-18 09:10:17","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":222815,"visible":true,"origin":"","legend":"\u003cp\u003eOverall architecture for clustering. The proposed framework extracts complementary features (RGB, HSV, GLCM, CLIP) from each image, concatenates them into a unified representation, and applies dimensionality reduction (T-SNE) followed by K-means clustering to group visually and semantically similar images.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-7315560/v1/b4f555cc64c33eca6dc2d377.png"},{"id":89273632,"identity":"0d7ee38f-5f0a-4c13-8ad0-f2618fafd920","added_by":"auto","created_at":"2025-08-18 09:10:17","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":348162,"visible":true,"origin":"","legend":"\u003cp\u003eClustering results of Korean modern and contemporary artists' paintings.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-7315560/v1/b9b31e7ac1ab0374eb64b05a.png"},{"id":89273633,"identity":"5ad02f81-3e82-4efb-a44a-6e0ee75e0a39","added_by":"auto","created_at":"2025-08-18 09:10:17","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":771879,"visible":true,"origin":"","legend":"\u003cp\u003eSchematic representation of representative images nearest to each cluster.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-7315560/v1/6b6453ddf853884018f6c737.png"},{"id":101690500,"identity":"8be94b91-7483-4b40-836c-1a94fad23922","added_by":"auto","created_at":"2026-02-02 16:04:19","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2030967,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7315560/v1/bd1671d5-00d9-4a97-859f-632e38980f46.pdf"},{"id":89273630,"identity":"3d62f391-aa00-4838-a121-3b32e345515f","added_by":"auto","created_at":"2025-08-18 09:10:17","extension":"doc","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":0,"visible":true,"origin":"","legend":"","description":"","filename":"Tables.doc","url":"https://assets-eu.researchsquare.com/files/rs-7315560/v1/574ac20ea9c5fbb6fe2382f8.doc"}],"financialInterests":"No competing interests reported.","formattedTitle":"Computational analysis of the 20th Century Korean paintings","fulltext":[{"header":"1\tIntroduction","content":"\u003cp\u003eAnalyzing artworks is inherently complex\u003csup\u003e1,2\u003c/sup\u003e. Art experts describe visual features such as space, texture, edges, form, shape, color, composition, lighting, brushstroke, tone, and line\u003csup\u003e3,4\u003c/sup\u003e. They also assess movement, harmony, balance, contrast, proportion, and pattern\u003csup\u003e5\u003c/sup\u003e. These elements provide quantifiable data and insights into an artist\u0026rsquo;s techniques, intentions, and narratives.\u003c/p\u003e\n\u003cp\u003eEach artwork carries a unique signature\u003csup\u003e6\u003c/sup\u003e. This distinctive visual style helps identify artistic connections and explain art movements and genres\u003csup\u003e7\u003c/sup\u003e. For instance, quantitative analysis of brushstroke configurations serves as a strong indicator of a painter\u0026rsquo;s style\u003csup\u003e8\u003c/sup\u003e. However, defining such styles can be challenging, as they often overlap, and artists may adopt multiple styles, complicating stylistic recognition\u003csup\u003e9,10\u003c/sup\u003e. For example, Pablo Picasso painted in both Surrealist and Cubist styles, continuously evolving over time. Moreover, human judgment in art is highly subjective and often controversial\u003csup\u003e11,12\u003c/sup\u003e, as experts rely heavily on personal experience and knowledge\u003csup\u003e13\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eThe collaboration between art and technology has a long history, with science supporting art experts since the early 18th century\u003csup\u003e14\u003c/sup\u003e. The development of computerized image processing techniques, including ultraviolet fluorescence, infrared reflectography, stereo microscopy, and X-radiography\u003csup\u003e15\u003c/sup\u003e, along with machine learning and computer vision algorithms\u003csup\u003e16\u003c/sup\u003e, has enabled advanced computational analysis as an interdisciplinary tool for examining paintings and deepening our understanding of art\u003csup\u003e17,18\u003c/sup\u003e. Additionally, the growing availability of large image datasets, such as WikiArt and ImageNet\u003csup\u003e19\u003c/sup\u003e, has further expanded the possibilities for computational art analysis\u003csup\u003e4,20\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eComputational algorithms are valuable for identifying and comparing artistic styles and similarities between paintings. Machine learning models can encode. discriminative visual features\u003csup\u003e21\u003c/sup\u003e and serve as effective tools for detecting forgeries and authenticating unknown artworks\u003csup\u003e1,21,22\u003c/sup\u003e. Additionally, deep neural networks can uncover hidden patterns, signatures, and meaningful relationships among artworks while classifying painting styles\u003csup\u003e9,12\u003c/sup\u003e. Studies suggest these computational methods often surpass even highly trained art experts in accuracy\u003csup\u003e23\u003c/sup\u003e, reinforcing computational aesthetics as a powerful approach for analyzing visual properties in the era of digital art history\u003csup\u003e24,25\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eResearch on AI-based art classification has taken various approaches, primarily utilizing Convolutional Neural Networks (CNN)\u003csup\u003e26-28\u003c/sup\u003e and Attention Mechanisms\u003csup\u003e29-31\u003c/sup\u003e. Traditional machine learning methods have also been employed, combining feature extraction techniques such as Principal Component Analysis (PCA)\u003csup\u003e32\u003c/sup\u003e or Histogram of Oriented Gradients (HOG)\u003csup\u003e33\u003c/sup\u003e with K-means clustering\u003csup\u003e34-36\u003c/sup\u003e. However, most of these studies focus on Western artworks\u003csup\u003e37\u003c/sup\u003e, leading to a bias toward specific art movements, while research on traditional Korean paintings remains relatively scarce.\u0026nbsp;\u003c/p\u003e"},{"header":"2\tRelated work","content":"\u003cp\u003eIn this study, we conduct an extensive literature review, identifying key trends and characteristics in the computational analysis of artworks. Additionally, we provide a comprehensive scholarly examination of representative modern and contemporary Korean painters, analyzing their distinctive styles and historical significance within the broader context of art history.\u003c/p\u003e\n\u003ch3\u003eClassification of art historical movements\u003c/h3\u003e\n\u003cp\u003eThe majority of studies on computational analysis have focused on the automatic classification of artworks based on categories such as artist, style, or genre. Several studies have specifically addressed automatic artist classification and identification\u003csup\u003e1,10\u003c/sup\u003e, style classification\u003csup\u003e5,9\u003c/sup\u003e, and genre classification\u003csup\u003e17,38\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003ePrevious studies on artist identification, artistic style, and art movement classification have extracted various features, including handcrafted features for classical machine learning\u003csup\u003e6,39\u003c/sup\u003e and deep features to characterize digitized paintings for deep learning\u003csup\u003e1,10,40\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eRegarding handcrafted features, some researchers have utilized low-level color and texture features\u003csup\u003e2,41,42\u003c/sup\u003e, while others have employed high-level semantic features\u003csup\u003e43\u003c/sup\u003e or a combination of both\u003csup\u003e5\u003c/sup\u003e. In studies of Vincent van Gogh\u0026rsquo;s paintings, quantitative features such as brushstroke distribution, orientation, width, length, color palette, composition, and shape have contributed to defining a unique signature of the artist\u0026rsquo;s style and aiding artist identification\u003csup\u003e14,41,44\u003c/sup\u003e.\u003c/p\u003e\n\u003ch3\u003eFocusing on classification models\u003c/h3\u003e\n\u003cp\u003eRecent classification studies have primarily used Generative or Discriminative models, or both\u003csup\u003e30,45,46\u003c/sup\u003e. Some focus exclusively on Generative models, while others rely solely on Discriminative models. Additionally, Graphical models have been explored for representation learning, and Hypothesis Matching models have been applied for data analysis.\u003c/p\u003e\n\u003cp\u003eThe classification of artworks using CNNs depends on extracting meaningful features while maintaining efficiency. Recent advancements in Attention Mechanisms allow models to capture detailed stylistic elements and distinguish subtle differences between art styles, addressing limitations in conventional CNN models\u003csup\u003e30,45,46\u003c/sup\u003e. By selectively emphasizing relevant image regions, attention-based approaches improve classification accuracy.\u003c/p\u003e\n\u003cp\u003eThe extraction of color and texture features is essential for distinguishing artistic styles, as different styles exhibit unique color schemes and brushwork. Efficiently integrating these features with neural networks enhances learning and classification accuracy in automated art analysis\u003csup\u003e47-49\u003c/sup\u003e. The Multi-Class Kernel Method improves classification robustness, especially for figurative styles with diverse feature components such as chromatic properties, textures, morphology, and composition. By mapping these features into high-dimensional spaces, this method captures complex, non-linear relationships, enabling more precise differentiation between artistic styles\u003csup\u003e50\u003c/sup\u003e. Additionally, transfer learning leverages pre-trained models to enhance performance on new tasks by transferring knowledge from large-scale datasets, reducing the need for extensive computational resources\u003csup\u003e45\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eMany researchers have used WikiArt\u003csup\u003e51\u003c/sup\u003e, one of the most comprehensive large-scale datasets, containing 150,000 artworks from 2,500 artists\u003csup\u003e8,10\u003c/sup\u003e. Other datasets include ArtCyclopedia\u003csup\u003e52\u003c/sup\u003e, Artstor Digital Library\u003csup\u003e43\u003c/sup\u003e, BBC Painting Dataset\u003csup\u003e53\u003c/sup\u003e, Mark Harden\u0026rsquo;s Artchive\u003csup\u003e5\u003c/sup\u003e, ABC Gallery\u003csup\u003e9\u003c/sup\u003e, and Artlex \u0026amp; CARLI Digital Collections\u003csup\u003e7\u003c/sup\u003e. Some studies also gathered images from Wikipedia, Flickr, online museums, and internet sources\u003csup\u003e54,55\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003ePark et al.\u003csup\u003e\u0026nbsp;56\u003c/sup\u003e selected images from 25 artists in the WikiArt dataset and addressed class imbalance using a weighted cross-entropy loss function. They expanded the dataset by applying augmentation techniques (e.g., resizing, flipping, rotation via Albumentations). CLAHE improved contrast, while CutMix enhanced texture information. By fine-tuning ResNet50, adjusting fully connected layers, and freezing convolutional weights, they significantly improved artist classification accuracy.\u003c/p\u003e\n\u003cp\u003ePrevious studies have explored the classification of traditional Chinese paintings (TCP)\u003csup\u003e57-59\u003c/sup\u003e. Li \u0026amp; Wang\u003csup\u003e57\u003c/sup\u003e used wavelets and 2D Multi-resolution Hidden Markov Models (MHMM) to categorize Chinese ink paintings by style and artist. Jiang et al.\u003csup\u003e58\u003c/sup\u003e distinguished TCP from non-TCP images and classified them into Gongbi (skilled brush, 1,889 images) and Xieyi styles using low-level features (color, texture, edge) and a hybrid classifier (decision tree + SVM). Their approach achieved practically viable accuracy. Lu et al.\u003csup\u003e59\u003c/sup\u003e developed a TCP classification framework based on four art movements (Xieyi, Gongbi, Goule, Shese) and six painters. They applied Bayesian classifiers, k-NN, fuzzy C-means clustering, and a non-linear multi-class SVM, comparing their classification performance.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eRecent studies have increasingly focused on improving artwork classification accuracy through data augmentation and advanced deep learning models. Baldrati et al.\u003csup\u003e60\u003c/sup\u003e introduced a CLIP-based multi-modal approach using the NoisyArt dataset, combining textual and visual features to enhance classification and retrieval tasks. This study demonstrated the effectiveness of multi-modal learning in computational art analysis. Zhong et al.\u003csup\u003e61\u003c/sup\u003e proposed a Two-Channel Dual-Path Network (FPTD) that incorporates RGB and brushstroke texture information to improve fine-art painting classification. The study used a Gray-Level Co-Occurrence Matrix (GLCM) to extract texture details from multiple directions, enabling more precise classification of styles, artists, and genres while enhancing generalization performance. Kim et al.\u003csup\u003e62\u003c/sup\u003e developed a proxy learning method that integrates pre-trained language models with visual data for analyzing artistic styles. By modeling the relationships between textual descriptions and visual features, their method extracts meaningful visual concepts, significantly improving automated classification and analysis of artworks. This interdisciplinary approach broadens traditional feature extraction techniques, offering new insights into computational art analysis.\u003c/p\u003e\n\u003ch3\u003eModern and contemporary Korean artists\u003c/h3\u003e\n\u003cp\u003eKim Ki-chang, Kim Whan-ki, Do Sang-bong, Park Soo-keun, Yoo Young-kuk, Lee Jung-seob, Chun Kyung-ja, Chang Uc-chin, Byun Kwan-sik, Lee Sang-beom, and Byun Jong-ha are key figures in modern and contemporary Korean art. Active in the 20th-century Korean art scene, they played a crucial role in shaping Korean modernism and establishing the national artistic identity in the global art world. Amidst the turbulent history of modern Korea, these artists developed distinctive artistic styles, producing works that reflect the evolution of Korean art. Among them, Kim Whan-ki, Park Soo-keun, Yoo Young-kuk, Lee Jung-seob, and Chang Uc-chin are regarded as the five leading second-generation Western-style painters and first-generation modernists in Korean art\u003csup\u003e63\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eKim Whan-ki (1913\u0026ndash;1974) and Yoo Young-kuk (1916\u0026ndash;2002) explored Korean modernity through experimental abstractions that bridged tradition and contemporary life. While Yoo Young-kuk pursued pure geometric abstraction, continuing pre-war abstract movements\u003csup\u003e64\u003c/sup\u003e, Kim Whan-ki integrated real-world motifs to express the spirit of the times\u003csup\u003e65\u003c/sup\u003e. Lee Jung-seob (1916\u0026ndash;1956), known for his restrained color palette, depicted local sentiments through motifs such as cows and children\u003csup\u003e66\u003c/sup\u003e. Chang Uc-chin (1917\u0026ndash;1990) explored formative art inspired by his hometown imagery. Park Soo-keun (1914\u0026ndash;1965) captured ordinary people\u0026rsquo;s daily lives during the Japanese colonial period and the Korean War, expressing their resilience and humanity with a \u0026ldquo;sincere heart and kind gaze\u0026rdquo;\u003csup\u003e\u0026nbsp;67\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eExcept for Park Soo-keun, the other four artists were members of the New Realism Group (Shinsasilpa), an influential painters\u0026apos; collective founded in July 1947\u003csup\u003e68\u003c/sup\u003e. Comprising mostly graduates of Japanese art schools, the group sought to merge modernist principles with traditional Korean identity, responding to the post-liberation and wartime political upheavals.\u003c/p\u003e\n\u003cp\u003eKim Ki-chang (1914\u0026ndash;2001) was an Oriental painter known for his \u0026quot;Foolish Painting Style,\u0026quot; which modernized traditional Joseon Dynasty folk and genre paintings\u003csup\u003e69\u003c/sup\u003e. Do Sang-bong (1902\u0026ndash;1977) integrated Korean sentiment into traditional realist painting from the late 1920s to the 1970s, employing balanced compositions and rich color palettes\u003csup\u003e63\u003c/sup\u003e. Chun Kyung-ja (1924\u0026ndash;2015), the only female artist among the eleven, forged a distinctive style by blending traditional Korean colors with bold, diverse hues. In an era dominated by ink wash painting, she introduced new possibilities for colored painting, merging Korean sentiment with modern aesthetics\u003csup\u003e70\u003c/sup\u003e. Byun Kwan-sik (1899\u0026ndash;1976), active from the 1920s to the mid-1970s, upheld the tradition of Korean painting while pioneering the \u0026quot;Sojeong style,\u0026quot; which featured distinctive compositions and varied ink techniques\u003csup\u003e71\u003c/sup\u003e. Lee Sang-beom (1897\u0026ndash;1972), influenced by photography and Western painting from 1923 onward, became a master of ink wash landscape painting. He developed the \u0026quot;Cheongjeon style,\u0026quot; shifting from conceptual landscapes to depictions of real scenery, incorporating mijeomjun techniques to modernize traditional landscape painting\u003csup\u003e72\u003c/sup\u003e. Byun Jong-ha (1926\u0026ndash;2000), active from the 1940s to 2000, developed a \u0026quot;three-dimensional painting\u0026quot; style in the 1960s, introducing depth and spatiality to flat paintings. By blurring the boundary between painting and sculpture, his work became a source of inspiration\u003csup\u003e73\u003c/sup\u003e.\u003c/p\u003e"},{"header":"3\tMethodology","content":"\u003cp\u003eImage datasets\u003c/p\u003e\n\u003cp\u003eIn this study, we constructed a dataset comprising 1,100 images from 11 representative modern and contemporary Korean painters, with 100 images per artist (Table 1). These were sourced primarily from the MMCA, which, as of October 28, 2024, holds 11,479 images, including 3,608 modern paintings. To expand the dataset, we gathered additional high-resolution images from reputable sources such as Google Arts \u0026amp; Culture and Google Search. We ensured authenticity by selecting images from authoritative websites that provided verified artwork details, including the title and year of creation.\u003c/p\u003e\n\u003cp\u003eUsing Google Search, we applied the \u0026apos;artist\u0026apos;s name\u0026apos; as a keyword and the \u0026apos;large size\u0026apos; filter to obtain high-resolution images. For missing metadata (e.g., title, creation year), we utilized Google Lens to perform reverse image searches, retrieving verified details from artist foundation websites and other credible sources. When low-resolution images were the dataset with verified artworks, ensuring data reliability and comprehensive coverage of modern and contemporary Korean paintings.\u003c/p\u003e\n\u003cp\u003eFramework\u003c/p\u003e\n\u003cp\u003eThis study proposes an analytical framework that integrates multiple image feature extraction techniques, dimensionality reduction, and clustering to analyze artworks by modern and contemporary Korean painters (Figure 1). The framework consists of the following steps: First, input data comprises images categorized by artist, where object regions within images are cropped based on pre-existing annotation data before processing. The collected artwork images are then transformed into diverse visual feature vectors using multiple image encoders. Specifically, four encoding methods\u0026mdash;RGB means\u003csup\u003e74\u003c/sup\u003e, HSV means\u003csup\u003e48\u003c/sup\u003e, color histograms, Gray-Level Co-occurrence Matrix (GLCM)\u003csup\u003e75\u003c/sup\u003e for texture analysis, and CLIP embeddings\u003csup\u003e76\u003c/sup\u003e\u0026mdash;are employed to extract distinct feature representations. Among these, RGB and HSV were employed as color spaces, while GLCM served as the texture feature space. These features are subsequently normalized and concatenated to construct a comprehensive multimodal feature vector for each image. Each feature vector is stored along with its corresponding filename for reference. The feature vectors of all images are then structured into a feature matrix, encapsulating the visual characteristics of the entire dataset. To facilitate visualization of high-dimensional features in a low-dimensional space, we apply t-SNE, while K-means clustering is used to identify typological similarities and group characteristics among the artwork images.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eUpon image input, feature vectors are generated through four modules: RGB, HSV, GLCM, and CLIP. The RGB module extracts color composition via red, green, and blue channels, while the HSV module computes mean and standard deviation values for each color component based on edge count, dark pixels, symmetry, and average values in hue, saturation, and value spaces. GLCM, a statistical texture analysis method, characterizes textures by quantifying the frequency of specific pixel pair occurrences at predefined spatial relationships, enabling detailed statistical measurements. Lastly, CLIP learns semantic associations between text and images, encoding input images into embedding vectors and linking them with textual descriptions. The model calculates cosine similarity between image and text embeddings, selecting the most relevant caption. This study employs pre-trained CLIP encoders for improved feature representation.\u003c/p\u003e\n\u003cp\u003eThe clustering module employs K-means clustering on low-dimensional feature vectors projected by t-SNE, automatically grouping modern and contemporary Korean paintings by type. Feature vectors from each module are concatenated into a unified vector per image, which is then vertically stacked across all images to form a feature matrix. This high-dimensional matrix is reduced to two dimensions\u0026mdash;x and y axes\u0026mdash;via t-SNE, enabling visual representation. By integrating the four modules, the matrix effectively captures key color and texture attributes, aiding in classification. The two-dimensional vectors, preserving essential visual features in a lower-dimensional space, serve as input for clustering.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eSubsequently, K-means clustering partitions the dataset into 11\u0026times;20 initial clusters. The K value was chosen to account for the diverse styles and evolving techniques of the 11 modern and contemporary Korean painters in the dataset. These initial clusters aim to capture fine stylistic variations, maximizing granularity in the early stages. Post-processing was applied to refine clustering by analyzing cluster sizes, removing smaller clusters, and retaining only the upper half to enhance interpretability. Final clustering was then performed by using the central coordinates of the retained clusters and reapplying K-means.\u003c/p\u003e\n\u003cp\u003eThe clustering results were evaluated by selecting representative images for each cluster and analyzing the distribution of ground truth labels (painter names) to assess cluster purity and average accuracy. The most frequently occurring ground truth label within each cluster was assigned as its representative label, quantifying how well clusters correspond to individual painters\u0026apos; stylistic tendencies. Classification outcomes were visualized using representative images and captions, providing interpretable insights into model performance. Additionally, clusters were displayed with circle colors indicating ground truth labels and outline colors representing clustering results, enabling intuitive verification of label alignment. This clustering strategy balances capturing artistic diversity in the initial stages while refining clusters for interpretability. Given that modern and contemporary Korean painters exhibit both distinctive personal styles and intra-artist variations across periods and subjects, this hierarchical clustering methodology effectively accommodates the dataset\u0026rsquo;s characteristics.\u003c/p\u003e"},{"header":"4\tExperiments","content":"\u003cp\u003eImplementation detail\u003c/p\u003e\n\u003cp\u003eIn this study, we constructed feature vectors by extracting image embeddings using the pre-trained CLIP model ViT-B/32 and integrating additional color and texture information. The experiments were conducted on an NVIDIA RTX 3090 GPU with Python 3.10, running on Ubuntu 22.04 with CUDA 12.2. Input images were normalized to 256\u0026times;256 resolution, and a batch size of 32 was set for CLIP embedding extraction.\u003c/p\u003e\n\u003cp\u003eFeature vector construction included one-dimensional RGB and HSV means per channel, 768-dimensional color histograms (256 per channel), and 23-dimensional GLCM features based on co-occurrence matrices with 23 gray levels and a 1-pixel offset. Additionally, 529-dimensional LBP features were generated using 23\u0026times;23 patches with a radius of 1, 8 neighbors, and bilinear resampling. These features were concatenated sequentially with CLIP embeddings, forming a final 1338-dimensional feature vector, with z-score normalization applied before clustering.\u003c/p\u003e\n\u003cp\u003eFor t-SNE-based dimensionality reduction, a learning rate of 200 and perplexity of 30 were set. K-Means clustering was initially performed with 20\u0026times; the number of artists, followed by filtering to retain the top 50% of clusters by size. Clustering accuracy was assessed based on the proportion of majority labels within each cluster. To ensure experimental reproducibility, random seeds were fixed for all t-SNE and K-Means processes.\u003c/p\u003e\n\u003cp\u003eResult and discussion\u003c/p\u003e\n\u003cp\u003eWe summarize the clustering results in Table 2, with the best results highlighted in bold. Data clustering was performed based on the visual characteristics of images, and the consistency within each cluster was assessed through intuitive visualization and quantitative evaluation. Specifically, visual features extracted from the pre-trained CLIP model were combined with RGB, HSV, histogram-based color features, and texture features derived from LBP (Local Binary Pattern)\u0026nbsp;and GLCM. These features were projected into a lower-dimensional space using t-SNE and clustered into 11 primary groups using the K-Means algorithm.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2.\u003c/strong\u003e Classification accuracy results by method for artist identification (unit: %).\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"463\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDataset\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"6\" valign=\"top\" style=\"width: 340px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMethod\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 47px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRGB\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 57px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eHSV\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 66px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eHistogram\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 57px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eLBP\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 57px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eGLCM\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 57px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eOurs\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 123px;\"\u003e\n \u003cp\u003eKorean modern painting\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 47px;\"\u003e\n \u003cp\u003e82.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 57px;\"\u003e\n \u003cp\u003e81.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 66px;\"\u003e\n \u003cp\u003e51.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 57px;\"\u003e\n \u003cp\u003e68.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 57px;\"\u003e\n \u003cp\u003e73.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 57px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e82.4\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eQuantitative performance evaluation was conducted by assigning the most frequently occurring label within each cluster as the representative label and defining accuracy as the proportion of images matching this label. This metric effectively quantifies cluster purity and is particularly suitable when ground truth labels are clearly defined (e.g., classification by artist or artwork type). Additionally, it enabled direct comparison between our method and previous feature extraction techniques. Prior studies primarily relied on RGB, HSV, and histogram-based features\u003csup\u003e77\u003c/sup\u003e for color extraction, while texture analysis was performed using LBP and GLCM to capture stylistic elements such as brushstroke patterns. Our method, which integrates these traditional features with a comprehensive visual-semantic representation, outperformed previous approaches, achieving an accuracy of 82.4% under the same dataset and experimental conditions. This result suggests that our approach more effectively captures the visual characteristics of artworks by complementing conventional color and texture features with richer semantic information.\u003c/p\u003e\n\u003cp\u003eWhile quantitative performance evaluation is standard in traditional machine learning research, qualitative interpretation often dominates studies in artistic data and visual cultural heritage. However, this study incorporates essential quantitative clustering accuracy measurements to ensure the objectivity of machine learning-based analysis and enhance comparability with future research. This approach goes beyond merely reporting classification accuracy; it serves as a validation process to determine whether each cluster is meaningfully structured around distinct visual characteristics.\u003c/p\u003e\n\u003cp\u003eThrough additional analysis, we compared the performance of traditional color-and texture-based methods with our proposed approach, categorized by artist and artwork type, as summarized in Table 3. Color-based models, such as RGB and HSV, exhibited high accuracy for artists known for their distinctive color usage. For example, Yoo Young-kuk, who frequently employed bold primary colors and geometric color planes, achieved 85.0% accuracy with RGB and 92.0% with HSV. Similarly, Kim Whan-ki recorded 72.0% and 75.0% accuracy in RGB and HSV, respectively. Do Sang-bong also demonstrated exceptionally high recognition rates in color-based analysis due to his consistent color schemes in still life paintings, characterized by recurring tonalities and specific color harmonies. These results suggest that color-based analytical methods are particularly effective when an artist\u0026rsquo;s signature color palette is a defining feature of their work.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 3.\u003c/strong\u003e Comparison of model accuracy by artist (unit: %).\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"463\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" valign=\"top\" style=\"width: 95px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDataset\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"6\" valign=\"top\" style=\"width: 369px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMethod\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRGB\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eHSV\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eHistogram\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eLBP\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eGLCM\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eOurs\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003eKim Ki Chang\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e41.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e44.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e33.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e34.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e42.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e73.5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003eKim Whan Ki\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e72.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e75.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e41.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e66.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e76.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e74.8\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003eTo Sang Bong\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e98.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e98.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e37.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e98.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e100\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e97.0\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003ePark Soo Keun\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e56.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e56.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e38.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e50.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e56.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e69.1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003eByeon Gwan Sik\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e50.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e49.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e30.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e53.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e51.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e76.0\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003eByun Jong Ha\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e49.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e58.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e24.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e36.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e51.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e78.3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003eYoo Young Kuk\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e85.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e92.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e56.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e70.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e93.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e94.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003eLee Sang Beom\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e65.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e64.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e32.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e63.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e60.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e78.4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003eLee Jung Seop\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e43.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e42.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e27.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e26.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e40.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e74.4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003eChang Uc Chin\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e49.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e47.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e24.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e36.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e60.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e86.4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003eCheon Gyeong Ja\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e67.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e75.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e18.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e65.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e74.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 61px;\"\u003e\n \u003cp\u003e82.3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eConversely, histogram-based methods showed relatively low accuracy across all cases, indicating that simple color distribution analysis alone is insufficient to capture the nuanced stylistic and painterly characteristics of each artist. Notably, Byun Jong-ha and Chang Uc-chin recorded accuracy rates below 30%, highlighting the limitations of relying solely on basic color frequency comparisons for artist identification.\u003c/p\u003e\n\u003cp\u003eThe LBP method, which learns patterns by analyzing image textures and tonal variations, demonstrated high performance for Do Sang-bong, likely due to the detailed brushwork and shading characteristic of his still life paintings. Yoo Young-kuk also showed strong performance, benefiting from the high contrast between color fields in his works. Meanwhile, the GLCM method, which captures patterns by analyzing spatial relationships between pixels, achieved 100% accuracy for Do Sang-bong, attributed to the consistent composition and recurring patterns in his still life paintings.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eIn contrast, artists such as Kim Ki-chang and Lee Jung-seob, whose works feature more spontaneous and dynamic compositions, exhibited relatively lower performance in pattern-based analysis. Similarly, the experimental and complex techniques of Byun Jong-ha, as well as the simplified and flat compositions characteristic of Chang Uc-chin, were not optimally captured by any specific method, leading to overall lower accuracy rates for these artists.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThese findings suggest that color-based techniques are more effective for artists who emphasize distinctive color usage, while texture and pattern analysis methods perform better for artists whose works focus on compositional structure and surface textures. This underscores the need for selecting analytical methods based on the inherent visual characteristics of each artist\u0026rsquo;s body of work.\u003c/p\u003e\n\u003cp\u003eIn summary, conventional methods demonstrated relatively high accuracy for certain artists but showed significant performance degradation for those employing distinctive or unconventional expressive techniques. This limitation stems from traditional methods\u0026apos; sensitivity to specific color combinations or surface texture patterns, which prevents them from fully capturing the holistic visual context of artworks. In contrast, the proposed method, which integrates CLIP-based features with color and texture information, achieved stable and consistent accuracy across all artists and artwork categories. By comprehensively reflecting diverse visual characteristics without bias toward specific artists or styles, this approach overcomes the limitations of conventional methods.\u003c/p\u003e\n\u003cp\u003eThis result is visualized in Figure 2. Using K-means clustering, the images were grouped into 11 main clusters, distinguished by color variations. Each artwork is represented as a point, with dark-bordered points indicating correctly classified images and lighter-bordered points representing misclassifications, allowing for easy verification of accuracy.\u003c/p\u003e\n\u003cp\u003eThe high proportion of dark-bordered points suggests that, despite being an unsupervised learning model, the system effectively identifies similarities among artworks, accurately classifying them by artist. This highlights the model\u0026rsquo;s strong ability to distinguish artists based solely on image information, underscoring the value of AI-driven analysis.\u003c/p\u003e\n\u003cp\u003eFigure 3 presents a schematic visualization of representative images closest to each cluster centroid. These images encapsulate the dominant visual characteristics of their respective clusters, which are grouped based on stylistic attributes such as color composition, texture patterns, and compositional structures. For instance, one cluster primarily consists of traditional ink paintings with monochromatic tonal variations and brushwork textures, while another is dominated by abstract works featuring bold primary colors and geometric forms. Still life paintings form a distinct cluster as well. This visualization confirms that the clustering process successfully captures the latent visual attributes of each artwork. By inspecting representative images within each cluster, it is evident that the proposed analytical approach effectively groups visually similar works, offering an interpretable mapping of stylistic relationships between artists and artworks.\u003c/p\u003e\n\u003cp\u003eAdditionally, to analyze differences between groups, the generated cluster centroids were examined. These centroids represent the average position of data points within each cluster, effectively summarizing the dominant color tones and texture characteristics of each group. Table 4 presents the centroids of correctly classified clusters, demonstrating that the generated captions accurately reflect the visual attributes of each group. This suggests that both the clustering and caption generation models exhibit high reliability in capturing distinctive visual features. However, as shown in Table 5, certain clusters produced incorrect classifications and captions, highlighting areas for further refinement.\u0026nbsp;\u003c/p\u003e"},{"header":"5\tConclusion","content":"\u003cp\u003eThis study introduces a novel analytical framework for analyzing and classifying artworks by modern and contemporary Korean painters, integrating color, texture, and semantic features. While previous research has largely focused on Western paintings, studies on Eastern art\u0026mdash;particularly Korean paintings\u0026mdash;remain limited. To address this gap, we constructed a dedicated dataset of works by major Korean artists and developed a methodology that captures their unique visual language and aesthetic characteristics.\u003c/p\u003e\n\u003cp\u003eComparative analysis of various feature extraction methods reveals that artists with distinctive color palettes achieve high accuracy in color-based analysis, while those emphasizing texture and brushwork perform better in texture-based approaches. In contrast, histogram-based methods relying solely on color frequency distributions struggle to capture unique artistic expressions. To overcome these limitations, this study presents an integrated approach that combines CLIP-based semantic features with visual attributes such as color and texture. The proposed method ensures stable and balanced performance across different artists and styles, effectively identifying visual similarities even in an unsupervised learning environment.\u003c/p\u003e\n\u003cp\u003eThe clustering results and representative image analysis confirm that the proposed framework effectively distinguishes artworks by artist. Representative images summarize each artist\u0026rsquo;s visual language in terms of color tone, texture, and composition, demonstrating that the framework not only enhances classification accuracy but also serves as a quantitative tool for identifying visual characteristics, exploring inter-artwork relationships, and situating works within broader art historical contexts.\u003c/p\u003e\n\u003cp\u003eAdditionally, this study contributes to the digital analysis of Korean modern and contemporary paintings by laying the groundwork for future artist classification models and stylistic analysis methods tailored to Korean art. It addresses the limitations of prior research, which predominantly focused on Western art while overlooking the unique visual culture and historical context of Korean paintings.\u003c/p\u003e\n\u003cp\u003eIn conclusion, this study introduces a novel AI-driven methodology that integrates visual analysis with semantic interpretation for the study of Korean modern and contemporary paintings. The experimental results highlight its effectiveness and academic significance, quantitatively capturing the distinct visual language and formal characteristics of Korean paintings. The dataset and analytical framework developed here offer a foundation for future research in digital art history. Expanding the dataset to include traditional Korean paintings and postmodern works would enable a more comprehensive chronological analysis, further illuminating the stylistic evolution and defining characteristics of Korean art across different periods and artists.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eAvailability of data\u003c/p\u003e\n\u003cp\u003eThe dataset used in this study consists of paintings by eleven modern and contemporary Korean artists. The images were provided by various institutions, including the National Museum of Modern and Contemporary Art (MMCA), the Jang Uc-Chin Foundation, Seoul Metropolitan Government, the Park Soo Keun Institute, and the Yoo Youngkuk Art Foundation. The copyright for the images belongs to the respective institutions.\u003c/p\u003e\n\u003cp\u003eCompeting interests\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003eAcknowledgements\u003c/p\u003e\n\u003cp\u003eThis research was supported by Culture, Sports and Tourism R\u0026amp;D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2023 (Project Name: Acquisition of 3D precise information of microstructure and development of authoring technology for ultra-high precision cultural restoration, Project Number: RS-2023-00227749, Contribution Rate: 100%)\u003c/p\u003e\n\u003cp\u003eAuthors\u0026apos; contributions\u003c/p\u003e\n\u003cp\u003eSH took primary responsibility for drafting the manuscript, designing the experiments, and conducting the main model experiments. SJ, SE, and YM contributed to data collection, data refinement, experimental interpretation, and the review of prior research. JW was responsible for the experimental design and interpretation of the results. BA supported the data refinement, validation of experimental results, and overall quality assurance. All authors read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003eAuthors\u0026apos; information\u003c/p\u003e\n\u003cp\u003eSeohyun Baek: First Author, Electronics and Telecommunications Research Institute (ETRI), Daejeon, Korea. E-mail: [email protected]\u003c/p\u003e\n\u003cp\u003eSo-Jeong Park: Contributing Author, Chung-Ang University, Seoul, Korea. E-mail: [email protected]\u003c/p\u003e\n\u003cp\u003eSo-Eun Park: Contributing Author, Chung-Ang University, Seoul, Korea. E-mail: [email protected]\u003c/p\u003e\n\u003cp\u003eYou-Min Im: Contributing Author, Chung-Ang University, Seoul, Korea. E-mail: [email protected]\u003c/p\u003e\n\u003cp\u003eJongwon Choi: Contributing Author, Chung-Ang University, Seoul, Korea. E-mail: [email protected]\u003c/p\u003e\n\u003cp\u003eBo-A Rhee: Corresponding Author, Chung-Ang University, Seoul, Korea. E-mail: [email protected]\u0026nbsp;\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eViswanathan, N. \u0026amp; Stanford. Artist identification with convolutional neural networks. https://api.semanticscholar.org/CorpusID:198974841 (2017).\u003c/li\u003e\n\u003cli\u003eKim, M. \u0026amp; Kim, J. Complementary quantitative approach to unsolved issues in art history: Similarity of visual features in the paintings of Vermeer and his probable mentors. \u003cem\u003eLeonardo\u003c/em\u003e \u003cstrong\u003e52\u003c/strong\u003e, 164-174 (2019).\u003c/li\u003e\n\u003cli\u003eShamir, L. et al. Impressionism, expressionism, surrealism: Automated recognition of painters and schools of art. \u003cem\u003eACM Transactions on Applied Perception\u003c/em\u003e \u003cstrong\u003e7\u003c/strong\u003e, 1-17 (2010).\u003c/li\u003e\n\u003cli\u003eCetinic, E., Lipić, T. \u0026amp; Grgic, S. Fine-tuning convolutional neural networks for fine art classification. Expert Syst. Appl 114, 107-118 (2018).\u003c/li\u003e\n\u003cli\u003eSaleh, B., Abe, K., Arora, R. \u0026amp; Elgammal, A. Toward automated discovery of artistic influence. \u003cem\u003eMultimedia Tools and Applications\u003c/em\u003e \u003cstrong\u003e75\u003c/strong\u003e, 3565-3591 (2016).\u003c/li\u003e\n\u003cli\u003eBar, Y., Levy, N. \u0026amp; Wolf, L. Classification of artistic styles using binarized features derived from a deep neural network. In \u003cem\u003eProc. ECCV Workshops\u003c/em\u003e, 71-84 (2014).\u003c/li\u003e\n\u003cli\u003eZujovic, J. et al. Classifying paintings by artistic genre: An analysis of features \u0026amp; classifiers. In \u003cem\u003eProc. IEEE international workshop on multimedia signal processing\u003c/em\u003e (2009).\u003c/li\u003e\n\u003cli\u003eSaleh, B. \u0026amp; Elgammal, A. A unified framework for painting classification. In \u003cem\u003eProc. IEEE International Conference on Data Mining Workshop\u003c/em\u003e (2015).\u003c/li\u003e\n\u003cli\u003eGultepe, E., Conturo, T. E. \u0026amp; Makrehchi, M. Predicting and grouping digitized paintings by style using unsupervised feature learning. \u003cem\u003eJournal of cultural heritage\u003c/em\u003e \u003cstrong\u003e31\u003c/strong\u003e, 13-23 (2018).\u003c/li\u003e\n\u003cli\u003eChen, J. \u0026amp; Deng, A. Comparison of machine learning techniques for artist identification. \u003cem\u003eStandford193CS231N Report\u003c/em\u003e; Art Computer Science: Stanford, CA, USA (2018).\u003c/li\u003e\n\u003cli\u003eBerezhnoy, I. E., Postma, E. O. \u0026amp; van den Herik, H. J. Authentic: computerized brushstroke analysis. In \u003cem\u003eProc. IEEE International Conference on Multimedia and Expo\u003c/em\u003e (2005).\u003c/li\u003e\n\u003cli\u003evan den Herik, H. J. \u0026amp; Postma, E. O. Discovering the visual signature of painters. In Future Directions for Intelligent Systems and Information Sciences: The Future of Speech and Image Technologies. Brain Computers, WWW, and Bioinformatics 129-147 (2000).\u003c/li\u003e\n\u003cli\u003eFalomir, Z., Lled\u0026oacute;, M., Sanz, I. \u0026amp; Gonzalez-Abril, L. Categorizing paintings in art styles based on qualitative color descriptors, quantitative global features and machine learning (QArt-Learn). \u003cem\u003eExpert Systems with Applications\u003c/em\u003e \u003cstrong\u003e97\u003c/strong\u003e, 83-94 (2018).\u003c/li\u003e\n\u003cli\u003eLamberti, F., Sanna, A. \u0026amp; Paravati, G. Computer-assisted analysis of painting brushstrokes: digital image processing for unsupervised extraction of visible features from van Gogh\u0026rsquo;s works. \u003cem\u003eEURASIP Journal on Image and Video Processing\u003c/em\u003e 1-17 (2014).\u003c/li\u003e\n\u003cli\u003eJohnson, R. C. et al. Image processing for artist identification. \u003cem\u003eIEEE Signal Processing Magazine\u003c/em\u003e \u003cstrong\u003e25\u003c/strong\u003e, 37-48 (2008).\u003c/li\u003e\n\u003cli\u003evan der Maaten, L. \u0026amp; Postma, E. Identifying the real van gogh with brushstroke textons. \u003cem\u003eWhite Paper\u003c/em\u003e\u003cem\u003e, Tilburg University\u003c/em\u003e (2009).\u003c/li\u003e\n\u003cli\u003eCondorovici, R. G., Vertan, C. \u0026amp; Florea, L. Artistic genre classification for digitized painting collections. \u003cem\u003eUPB Scientific Bulletin\u003c/em\u003e \u003cstrong\u003e75\u003c/strong\u003e, 75-86 (2013).\u003c/li\u003e\n\u003cli\u003eStork, D., Coddington, J. \u0026amp; Bentkowska-Kafel, A. Computer vision and image analysis of art. In \u003cem\u003eProceedings of SPIE\u003c/em\u003e \u003cstrong\u003e7531\u003c/strong\u003e (2010).\u003c/li\u003e\n\u003cli\u003eDeng, J. et al. Imagenet: A large-scale hierarchical image database. In \u003cem\u003eProc. IEEE conference on computer vision and pattern recognition\u003c/em\u003e (2009).\u003c/li\u003e\n\u003cli\u003eSigaki, H. Y. D., Perc, M. \u0026amp; Haroldo, R. V. History of art paintings through the lens of entropy and complexity. \u003cem\u003eNational Academy of Sciences\u003c/em\u003e \u003cstrong\u003e115\u003c/strong\u003e, E8585\u0026ndash;E8594 (2018).\u003c/li\u003e\n\u003cli\u003eElgammal, A. et al. The shape of art history in the eyes of the machine. In \u003cem\u003eProc. AAAI Conference on Artificial Intelligence\u003c/em\u003e \u003cstrong\u003e32\u003c/strong\u003e, 2183-2191 (2018).\u003c/li\u003e\n\u003cli\u003eNemade, R., Nitsure, A., Hirve, P. \u0026amp; Mane, S. B. Detection of forgery in art paintings using machine learning. \u003cem\u003eInt. J. Innov. Res. Sci. Eng. Technol\u003c/em\u003e \u003cstrong\u003e6\u003c/strong\u003e, 8681-8692 (2017).\u003c/li\u003e\n\u003cli\u003eStork, D. Computer image analysis of paintings and drawings: an introduction to the literature. In \u003cem\u003eProc. Image Processing for Artist Identification Workshop\u003c/em\u003e (2008).\u003c/li\u003e\n\u003cli\u003eCetinic, E., Tomislav, L. \u0026amp; Grgic, S. A deep learning perspective on beauty, sentiment, and remembrance of art. \u003cem\u003eIEEE Access\u003c/em\u003e \u003cstrong\u003e7\u003c/strong\u003e, 73694-73710 (2019).\u003c/li\u003e\n\u003cli\u003eEllis, M. H. \u0026amp; Johnson Jr., C. R. Computational connoisseurship: enhanced examination using automated image analysis. \u003cem\u003eVisual Resources\u003c/em\u003e \u003cstrong\u003e35\u003c/strong\u003e, 125-140 (2019).\u003c/li\u003e\n\u003cli\u003eYu, Q. \u0026amp; Shi, C. An image classification approach for painting using improved convolutional neural algorithm. Soft Computing 28, 847-873 (2024).\u003c/li\u003e\n\u003cli\u003eMondal, K. \u0026amp; Anita, H. Categorization of artwork images based on painters using CNN. \u003cem\u003eJ. Phys.: Conf. Ser.\u003c/em\u003e (2021).\u003c/li\u003e\n\u003cli\u003eCervantes, F. et al. Evaluation of CNN models with transfer learning in art media classification in terms of accuracy and class relationship. Computaci\u0026oacute;n y Sistemas 28, 233-244 (2024).\u003c/li\u003e\n\u003cli\u003eZhang, X. \u0026amp; Ding, T. Style classification of media painting images by integrating ResNet and attention mechanism. \u003cem\u003eHeliyon\u003c/em\u003e \u003cstrong\u003e10\u003c/strong\u003e, e27178 (2024).\u003c/li\u003e\n\u003cli\u003eLi, Q., Yue, L. \u0026amp; Feng, Q. Research on automatic classification method of artistic styles based on attention mechanism convolutional neural network. \u003cem\u003eInt. J. Low-Carbon Technol.\u003c/em\u003e \u003cstrong\u003e19\u003c/strong\u003e, 2018-2023 (2024).\u003c/li\u003e\n\u003cli\u003eZeng, L. Improved painting image style classification of ResNet based on attention mechanism. In \u003cem\u003eProc. International Conference on Advances in Electrical Engineering and Computer Applications\u003c/em\u003e (2023).\u003c/li\u003e\n\u003cli\u003ePolak, A. et al. Hyperspectral imaging combined with data classification techniques as an aid for artwork authentication. \u003cem\u003eJ. cultural heritage\u003c/em\u003e \u003cstrong\u003e26\u003c/strong\u003e, 1-11 (2017).\u003c/li\u003e\n\u003cli\u003eSahu, T., Tyagi, A., Kumar, S. \u0026amp; Mittal, A. Classification and aesthetic evaluation of paintings and artworks. In \u003cem\u003eProc. 13th International Conference on Signal-Image Technology \u0026amp; Internet-Based Systems\u003c/em\u003e (2017).\u003c/li\u003e\n\u003cli\u003eCastellano, G. \u0026amp; Vessio, G. A deep learning approach to clustering visual arts. \u003cem\u003eInt. J. Comput. Vis.\u003c/em\u003e \u003cstrong\u003e130\u003c/strong\u003e, 2590-2605 (2022).\u003c/li\u003e\n\u003cli\u003eAbas, F. S. \u0026amp; Martinez, K. Classification of painting cracks for content-based analysis. In \u003cem\u003eMachine Vision Applications in Industrial Inspection XI\u003c/em\u003e \u003cstrong\u003e5011\u003c/strong\u003e, 149-160 (SPIE, 2003).\u003c/li\u003e\n\u003cli\u003eAndronache, I. Can complexity measures with hierarchical cluster analysis identify overpainted artwork? \u003cem\u003eZenodo\u003c/em\u003e (2023).\u003c/li\u003e\n\u003cli\u003eBaek, S. et al. Identification of Korean Neo-Realism artists through CLIP-based analysis. \u003cem\u003eJ. Korea Soc. Comput. Inf.\u003c/em\u003e \u003cstrong\u003e29\u003c/strong\u003e, 317-328 (2024).\u003c/li\u003e\n\u003cli\u003eAgarwal, S., Karnick, H., Pant, N. \u0026amp; Patel, U. Genre and style based painting classification. In \u003cem\u003eProc. IEEE Winter Conference on Applications of Computer Vision\u003c/em\u003e (2015).\u003c/li\u003e\n\u003cli\u003eMensink, T. \u0026amp; Van Gemert, J. The rijksmuseum challenge: Museum-centered visual recognition. In \u003cem\u003eProc. International conference on multimedia retrieval\u003c/em\u003e (2014).\u003c/li\u003e\n\u003cli\u003eKim, D., Xu, J., Elgammal, A. \u0026amp; Mazzone, M. Computational analysis of content in fine art paintings. In \u003cem\u003eProc. ICCC\u003c/em\u003e (2019).\u003c/li\u003e\n\u003cli\u003eLi, J., Yao, L., Hendriks, E. \u0026amp; Wang, J. Z. Rhythmic brushstrokes distinguish van Gogh from his contemporaries: findings via automated brushstroke extraction. \u003cem\u003eIEEE Trans. Pattern Anal. Mach. Intell.\u003c/em\u003e \u003cstrong\u003e34\u003c/strong\u003e, 1159-1176 (2011).\u003c/li\u003e\n\u003cli\u003ePark, J. W. Secrets of balanced composition as seen through a painter\u0026apos;s window: visual analyses of paintings based on subset barycenter patterns. In \u003cem\u003eACM SIGGRAPH 2019 Art Gallery\u003c/em\u003e 1-10 (2019).\u003c/li\u003e\n\u003cli\u003eCarneiro, G. Graph-based methods for the automatic annotation and retrieval of art prints. In \u003cem\u003eProc. 1st ACM International conference on multimedia retrieval\u003c/em\u003e (2011).\u003c/li\u003e\n\u003cli\u003eDasgupta, S. \u0026amp; Hossain, M. A. Pattern recognition through brushstroke analysis: an inspection to classify van Gogh from others. \u003cem\u003eBRAC University\u003c/em\u003e, (2017).\u003c/li\u003e\n\u003cli\u003eXu, K. Automated artist identification using deep learning and transfer learning techniques. \u003cem\u003eSci. Technol. Eng.\u003c/em\u003e \u003cstrong\u003e1\u003c/strong\u003e, 8 (2024).\u003c/li\u003e\n\u003cli\u003eZheng, S. Deep learning-based artistic style transformation algorithm in visual communication. \u003cem\u003eInt. J. Syst. Assurance Eng. Manag.\u003c/em\u003e 1-10 (2024).\u003c/li\u003e\n\u003cli\u003eWei, N. Research on the algorithm of painting image style feature extraction based on intelligent vision. \u003cem\u003eFutur. Gener. Comput. Syst.\u003c/em\u003e \u003cstrong\u003e123\u003c/strong\u003e, 196-200 (2021).\u003c/li\u003e\n\u003cli\u003eNunez-Garcia, I. et al. Classification of paintings by artistic style using color and texture features. \u003cem\u003eComput. Sist.\u003c/em\u003e \u003cstrong\u003e26\u003c/strong\u003e, 1503-1514 (2022).\u003c/li\u003e\n\u003cli\u003eHu, B. \u0026amp; Yang, Y. Construction of a painting image classification model based on AI stroke feature extraction. \u003cem\u003eJ. Intell. Syst.\u003c/em\u003e \u003cstrong\u003e33\u003c/strong\u003e, 20240042 (2024).\u003c/li\u003e\n\u003cli\u003eWang, Q. Painting style recognition and classification based on convolutional neural network and multi-kernel learning. In \u003cem\u003eIntelligent Computing Technology and Automation\u003c/em\u003e 1059-1066 (2024).\u003c/li\u003e\n\u003cli\u003eWikiArt. \u003cem\u003eWikiArt Visual Art Encyclopedia\u003c/em\u003e. n.d. Web. 3 July 2025. http://www.wikiart.org.\u003c/li\u003e\n\u003cli\u003eIvanova, K. et al. Features for art painting classification based on vector quantization of mpeg-7 descriptors. In \u003cem\u003eProc. Data Engineering and Management: Second International Conference\u003c/em\u003e 178-189 (2012). \u003c/li\u003e\n\u003cli\u003eCrowley, E. J. \u0026amp; Zisserman, A. In Search of Art. In Agapito, L., Bronstein, M. M. \u0026amp; Rother, C. (eds.) \u003cem\u003eComputer Vision - ECCV 2014 Workshops: Zurich, Switzerland, September 6-7 and 12, 2014, Proceedings, Part I\u003c/em\u003e 54-70 (2015).\u003c/li\u003e\n\u003cli\u003eShamir, L. Computer analysis reveals similarities between the artistic styles of Van Gogh and Pollock. \u003cem\u003eLeonardo\u003c/em\u003e \u003cstrong\u003e45\u003c/strong\u003e, 149-154 (2012).\u003c/li\u003e\n\u003cli\u003eKarayev, S. et al. Recognizing image style. \u003cem\u003earXiv preprint arXiv:1311.3715\u003c/em\u003e (2013).\u003c/li\u003e\n\u003cli\u003ePark, J.S., Kim, S.Y., Yoon, Y.C. \u0026amp; Kim, S. K. Optimizing CNN structure to improve accuracy of artwork artist classification. \u003cem\u003eJ. Korea Soc. Comput. Inf.\u003c/em\u003e \u003cstrong\u003e28\u003c/strong\u003e, 9-15 (2023).\u003c/li\u003e\n\u003cli\u003eLi, J. \u0026amp; Wang, J. Z. Studying digital imagery of ancient paintings by mixtures of stochastic models. \u003cem\u003eIEEE Trans. Image Process.\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e, 340\u0026ndash;353 (2004).\u003c/li\u003e\n\u003cli\u003eJiang, S., Huang, Q., Ye, Q. \u0026amp; Gao, W. An effective method to detect and categorize digitized traditional Chinese paintings. \u003cem\u003ePattern Recognit. Lett.\u003c/em\u003e \u003cstrong\u003e27\u003c/strong\u003e, 734\u0026ndash;746 (2006).\u003c/li\u003e\n\u003cli\u003eLu, G. et al. Content-based identifying and classifying traditional chinese painting images. In \u003cem\u003eProc. Congress on Image and Signal Processing\u003c/em\u003e (2008).\u003c/li\u003e\n\u003cli\u003eBaldrati, A. et al. Exploiting CLIP-based multi-modal approach for artwork classification and retrieval. \u003cem\u003earXiv preprint arXiv:2309.12110\u003c/em\u003e (2023).\u003c/li\u003e\n\u003cli\u003eZhong, S., Huang, X. \u0026amp; Xiao, Z. Fine-art painting classification via two-channel dual path networks. \u003cem\u003eInt. J. Mach. Learn. Cybern.\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, 137\u0026ndash;152 (2020).\u003c/li\u003e\n\u003cli\u003eKim, D., Elgammal, A. \u0026amp; Mazzone, M. Formal analysis of art: proxy learning of visual concepts from style through language models. \u003cem\u003earXiv preprint arXiv:2201.01819\u003c/em\u003e (2022).\u003c/li\u003e\n\u003cli\u003eNational Museum of Modern and Contemporary Art, Korea. \u003cem\u003eThe Most Honest Confession: Chang Ucchin Retrospective\u003c/em\u003e. n.d. Web. 14 Sept. 2023. https://www.mmca.go.kr/exhibitions/exhibitionsDetail.do?menuId=1030000000\u0026amp;exhId=202302150001627.\u003c/li\u003e\n\u003cli\u003eJoo, S. A Study on the Sublime in YOO YOUNGKUK\u0026apos;s Abstract Paintings: Focusing on works since the 1960s. \u003cem\u003eDepartment of Aesthetics \u0026amp; Art History, Graduate School of Chosun University, Gwangju, Korea\u003c/em\u003e (2023).\u003c/li\u003e\n\u003cli\u003eKim, H.S. Study on the Colors of Kim Whan-ki\u0026apos;s Painting. \u003cem\u003eJ. Art Theory Pract.\u003c/em\u003e \u003cstrong\u003e3\u003c/strong\u003e, 155\u0026ndash;172 (2005).\u003c/li\u003e\n\u003cli\u003eKim, M.J. About Originality on Materials and Techniques in Lee Jung Seob\u0026rsquo;s Painting. \u003cem\u003eJ. Korean Mod. Contemp. Art Hist.\u003c/em\u003e \u003cstrong\u003e32\u003c/strong\u003e, 83\u0026ndash;117 (2016).\u003c/li\u003e\n\u003cli\u003eEom, S.M. A Unique Artistic World: Park Soo-keun\u0026rsquo;s Concave and Convex Technique. One Art World 30\u0026ndash;37 (2022).\u003c/li\u003e\n\u003cli\u003ePark, J. M. The concept of abstraction in New Realism group: New interpretation on geometric abstraction from its own point of view. \u003cem\u003eKorean Bulletin of Art History\u003c/em\u003e \u003cstrong\u003e38\u003c/strong\u003e, 240\u0026ndash;276 (2012).\u003c/li\u003e\n\u003cli\u003eChoi, B. The changing characteristics and meaning of Woonbo Kim, Ki-Chang\u0026rsquo;s \u0026lsquo;The Foolish Painting Style\u0026rsquo;. \u003cem\u003eOriental Art\u003c/em\u003e \u003cstrong\u003e33\u003c/strong\u003e, 233\u0026ndash;256 (2016).\u003c/li\u003e\n\u003cli\u003eLee, M.S. A Study on the Cheon, Kyung-Ja\u0026apos;s art works. \u003cem\u003eGraduate School of Jeju University, Jeju, Korea\u003c/em\u003e\u003cem\u003e. \u003c/em\u003e(2011).\u003c/li\u003e\n\u003cli\u003eLee, Y. U. A study on Byeon kwan sik\u0026rsquo;s Mt. Geumgang paintings. \u003cem\u003eArt Hist. Cult. Herit.\u003c/em\u003e \u003cstrong\u003e1\u003c/strong\u003e, 9\u0026ndash;34 (2012).\u003c/li\u003e\n\u003cli\u003eNational Museum of Modern and Contemporary Art, Korea. \u003cem\u003e\u0026ldquo;Today, This Work: Lee Sangbeom | Early Winter | 1926.\u0026rdquo;\u003c/em\u003e n.d. Web. 1 June 2023. https://www.mmca.go.kr/digitals/digitalMovInfo.do?mbId=202306010001268.\u003c/li\u003e\n\u003cli\u003eHur, N.Y. The Sensibility Narrative through figuration by Byun Jong-Ha. \u003cem\u003eAssoc. Biogr. Art Hist. Korea\u003c/em\u003e \u003cstrong\u003e5\u003c/strong\u003e, 265\u0026ndash;287 (2009).\u003c/li\u003e\n\u003cli\u003eMaxwell, J. C. Experiments on colour as perceived by the eye, with remarks on colour-blindness. \u003cem\u003eRoyal Society of Edinburgh\u003c/em\u003e \u003cstrong\u003e3\u003c/strong\u003e, 299\u0026ndash;301 (1857).\u003c/li\u003e\n\u003cli\u003eSiqueira, D., Roberti, F., Schwartz, W. R. \u0026amp; Pedrini, H. Multi-scale gray level co-occurrence matrices for texture description. \u003cem\u003eNeurocomputing\u003c/em\u003e \u003cstrong\u003e120\u003c/strong\u003e, 336\u0026ndash;345 (2013).\u003c/li\u003e\n\u003cli\u003eNukrai, D., Mokady, R. \u0026amp; Globerson, A. Text-only training for image captioning using noise-injected clip. \u003cem\u003earXiv preprint arXiv:2211.00575\u003c/em\u003e (2022).\u003c/li\u003e\n\u003cli\u003e\u0026Ouml;zl\u0026uuml;, Ahmet. \u003cem\u003eColor Recognition\u003c/em\u003e. GitHub, n.d. Web. 3 July 2024. https://github.com/ahmetozlu/color_recognition.\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003eTables 1 and 4 are available in the Supplementary Files section.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"npj-heritage-science","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"hsci","sideBox":"Learn more about [Heritage Science](http://heritagesciencejournal.springeropen.com)","snPcode":"40494","submissionUrl":"https://submission.nature.com/new-submission/40494/3","title":"npj Heritage Science","twitterHandle":"@SpringerOpen","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Artist identification, Computational aesthetics, Modern and contemporary paintings, Painting feature extraction, Unsupervised learning","lastPublishedDoi":"10.21203/rs.3.rs-7315560/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7315560/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"This study presents a machine learning-based approach for analyzing and classifying modern and contemporary Korean paintings using image data. The model leverages a visual-language multimodal model for efficient visual feature extraction and employs a multi-layered image analysis to capture detailed formal characteristics. Color features are extracted through analyses of various color spaces, while texture information is quantified within the texture feature space. The extracted feature vectors are analyzed and visualized through clustering, achieving an artist classification accuracy of 82.4%. Representative images from each artist cluster effectively encapsulate and highlight distinctive color and textural characteristics. Additionally, image captioning techniques were applied to generate textual descriptions of the representative images, successfully translating visual features into descriptive text. The findings confirm that machine learning-based image analysis offers an effective and objective methodology for identifying and classifying the unique characteristics of modern and contemporary Korean paintings.","manuscriptTitle":"Computational analysis of the 20th Century Korean paintings","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-08-18 09:10:12","doi":"10.21203/rs.3.rs-7315560/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-09-15T00:07:29+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-09-01T18:25:18+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-09-01T07:51:54+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-08-29T08:01:55+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-08-28T03:26:33+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-08-24T09:39:23+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-08-17T11:28:01+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"132562738228351851267599665627736306063","date":"2025-08-17T11:19:01+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-08-16T07:42:02+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"207937253039904011657098564467906446401","date":"2025-08-16T03:33:22+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"272272382756704667646565781155332419350","date":"2025-08-14T05:53:07+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"184842642857879151098642140395275771637","date":"2025-08-13T08:31:08+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"143715811718669678808912201442878400490","date":"2025-08-13T05:23:09+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"113437058957983168379440888360563360853","date":"2025-08-11T14:01:06+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"72940192677757367527780587692873141318","date":"2025-08-11T07:31:04+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"233731766952689903635189157404434864589","date":"2025-08-11T07:14:28+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"175937681743213703175600620886544111670","date":"2025-08-11T03:10:52+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-08-11T02:20:33+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-08-08T04:57:12+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-08-08T04:55:55+00:00","index":"","fulltext":""},{"type":"submitted","content":"npj heritage science","date":"2025-08-07T06:42:36+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"npj-heritage-science","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"hsci","sideBox":"Learn more about [Heritage Science](http://heritagesciencejournal.springeropen.com)","snPcode":"40494","submissionUrl":"https://submission.nature.com/new-submission/40494/3","title":"npj Heritage Science","twitterHandle":"@SpringerOpen","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"eddf9112-c7e8-4461-a678-21aa4bbc1afb","owner":[],"postedDate":"August 18th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2026-02-02T16:01:22+00:00","versionOfRecord":{"articleIdentity":"rs-7315560","link":"https://doi.org/10.1038/s40494-026-02304-1","journal":{"identity":"npj-heritage-science","isVorOnly":false,"title":"npj Heritage Science"},"publishedOn":"2026-01-31 15:58:45","publishedOnDateReadable":"January 31st, 2026"},"versionCreatedAt":"2025-08-18 09:10:12","video":"","vorDoi":"10.1038/s40494-026-02304-1","vorDoiUrl":"https://doi.org/10.1038/s40494-026-02304-1","workflowStages":[]},"version":"v1","identity":"rs-7315560","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7315560","identity":"rs-7315560","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}