Sketch-Based Manga Retrieval Using Deep Features
Rei Narita, Koki Tsubota, Toshihiko Yamasaki, Kiyoharu Aizawa · 2017
Manga, Japanese comics, are globally popular, and the digital manga (e-manga) market is growing year by year. E-manga has a limitation in its search methodology: it is currently restricted to a keyword search of authors and titles. In this paper, we present an intuitive sketch-based manga retrieval method using deep features. We propose a framework to extract feature vectors from sketches and manga images by two differently trained CNNs: The two CNNs are trained on a large number of manga face images with and without screentone. The deep features are switched according to a query, that is, a sketch drawing or a crop of a manga image. We built an interactive retrieval system that has a browser interface. We evaluated its retrieval accuracy by using sketches and manga images. The proposed method significantly outperformed state-of-the art sketch-based manga retrieval using handcrafted features.