A study on similarity-based clustering of manga character facial images toward advanced manga analysis

Abstract (English) The prevalence of pirated manga and unauthorized manga videos, infringing on intellectual property rights, is increasing significantly. To address this issue, we aim to develop methods for preventing manga piracy by identifying key information such as the author, work title, volume number, and episode number from manga page images, as well as by classifying character images. Our previous work focused on character face part extraction from manga page images. Using the Manga109 dataset, we trained a machine learning model to extract face part and evaluated its accuracy in identifying character face parts. In this paper, we attempt to cluster characters that appear in unknown manga page images using the face part extraction model which we created. We apply these models to manga page images, and the extracted face images are clustered using unsupervised learning to separate characters. This study contributes to advancing automated character analysis and character-based identification in manga images.