CLIP 기반 분석을 통한 한국 신사실파 작가 분류

초록

This study proposes a methodology for analyzing artwork features and classifying artists using image data. Unlike prior research focused on Western artworks, it builds and utilizes a dataset of Korean Neo-Realism (i.e. Shinsasilpa) artists. The image encoder of the CLIP model transformed the semantic features of the artwork into a vector space and, by learning together with the textual description, more deeply understood the meaning of the image. Furthermore, hue information via RGB and HSV color space analysis, and texture characteristics through GLCM-based analysis. These features were integrated into representative feature vectors and analyzed with K-means clustering, achieving 87.4% classification accuracy. Visualization results demonstrated the model's effectiveness in identifying image similarities and accurately classifying artworks in an unsupervised learning context, while highlighting unique hue and texture characteristics of each cluster, revealing formal and artistic tendencies in artworks.

키워드