Automatic Classifiaction of Manga Characters using Density-Based Clustering

EasyChair Preprint 2084

Abstract

Manga (Japanese comics) is a popular content worldwide. By extracting metadata from manga, it can be used to provide e-comic services. However, since comics have unique image features, special image recognition methods are required. Character is an important component for understanding the stories of manga. In order to extract character information with lower cost, a system that automatically classifies character images is required. In our existing research, we proposed automatic classification of character images using DBSCAN which is a clustering method based on data density. However, there is a problem that DBSCAN strongly depends on the hyperparameter setting. In this paper, we examined the application of other density-based clustering methods to simplify character classification. We also verified the changing of clustering results caused by different CNN model for image feature extraction.

Keyphrases: CNN, Clustering, HDBSCAN, Manga, optics

@booklet{EasyChair:2084,
  author    = {Hideaki Yanagisawa and Kengo Kyogoku and Ravi Jain and Hiroshi Watanabe},
  title     = {Automatic Classifiaction of Manga Characters using Density-Based Clustering},
  howpublished = {EasyChair Preprint 2084},
  year      = {EasyChair, 2019}}