On Style-based Manga Analysis: Feature Design and Style Modeling

Abstract

In this talk, we will describe style-based manga analyses from the perspectives of feature design and model construction. To describe manga pages, three types of manga-specific features are proposed, including line features showing strokes and characteristics of edges, screentone features showing texture and shade, and panel features showing panel arrangement. Based on these features, the latent Dirichlet allocation technique is utilized to discover latent style elements embedded in manga documents. Distributions of style elements are then used to measure similarity between manga documents, and facilitate the development of various style-based applications, such as style-based artist retrieval, art movement retrieval, and artwork period classification. Statistical analysis of the proposed features demonstrate effectiveness of the manga description, and various experimental results show that the style models are able to yield promising performance in many applications and could bring many potential extensions.