A Deep Learning Approach to Quantitative Analysis of Traditional Japanese Artworks: Exploring the Latent Features of Ukiyo-e

Abstract

This study quantitatively investigates ukiyo-e, a traditional Japanese artwork, and extracts the latent features of artworks using a deep learning model. We employ a deep learning model to extract the latent features of more than 9,000 ukiyo-e artworks, utilizing the VGG models and examining the model's layers to identify the characteristics of representative ukiyo-e artists. With the obtained model, we conduct a within-artist study to capture the transition of artists’ styles of each artist to understand the career trajectory of Ukiyo-e artists according to their school or popularity. While most previous studies focus on the part of the artwork, such as the faces of the portrait, our study aims to deliver a holistic study of ukiyo-e artwork based on the whole of the images, including backgrounds. Through this holistic study, we aim to advance the quantitative measurement of traditional artworks in Japan.