We are delighted to announce that two papers from the Visual Intelligence Lab (VI Lab, Prof. Jongwon Choi) have been accepted to npj Heritage Science.

Title:
Explainable machine learning-based classifi cation of traditional Korean ceramics using XRF chemical composition data

Authors:
Ye Eun Cho, Sona Sim, Jongwon Choi†, Sangdoo Ahn†

Abstract:
This study presents an explainable machine learning approach for classifying traditional Korean ceramics, including celadon, buncheong, and white porcelain, based on X-ray fluorescence chemical composition data. A curated dataset of 624 samples was analyzed using six machine learning algorithms: principal component analysis-linear discriminant analysis, decision tree, random forest, extreme gradient boosting, k-nearest neighbors, and support vector machine. Among them, tree-based random forest and extreme gradient boosting models achieved the highest classification accuracy of 95.8%. While white porcelain was accurately identified across all models, celadon and buncheong showed partial misclassification due to overlapping chemical characteristics. Model interpretability was enhanced using Shapley additive explanations, which identified Fe2O3 and TiO2 as the most influential components for type differentiation, consistent with established ceramic coloration mechanisms. These results demonstrate the effectiveness of explainable machine learning for chemical-based ceramic classification and provide a quantitative framework that complements traditional typological approaches in cultural heritage research.

--

Title:
Toward enhanced unsupervised clustering of 20th century Korean paintings via multimodal features

Authors:
Seohyun Baek, So-Jeong Park, So-Eun Park, You-Min Im, Jongwon Choi, Bo-A Rhee

Abstract:
This study presents a machine learning framework for analyzing and clustering modern and contemporary Korean paintings based on image data. A pretrained vision–language architecture combined with multi-layered analysis was used to efficiently extract detailed formal characteristics, including color features from multiple spaces and quantified texture. The extracted feature vectors are clustered and evaluated under majority-label assignment, achieving 82.4% overall accuracy, outperforming single-feature baselines (RGB 82.0%, HSV 81.3%, histogram 51.0%, LBP 68.8%, and GLCM 73.7%). The proposed method achieves higher per-artist precision, better boundary-case discrimination, and greater robustness for low-sample categories. Representative images from artist clusters encapsulate unique color and texture. Analysis was extended using automatic image captioning and zero-shot style assignment via matching image and text embeddings. The findings demonstrate that machine learning–based image analysis provides an effective and objective methodology for identifying and distinguishing visual characteristics of modern and contemporary Korean paintings, offering a quantitative approach to art-historical interpretation.