Similarity Learning Based on Pool-Based Active Learning for Manga Character Retrieval

要約

Manga is Japanese style comics which is one of the most popular publications in Japan. There are some e-book services for selling manga. Unfortunately, most of them handle manga only as the set of images. Some digital media use tags or text data for retrieval, where the tags and text data are produced by handmade input. Our final goal is developing a practical content-based manga retrieval system based on automatic tagging of the contents of manga, for example, place, character, expression, situation, and other information readers are interested in. At the firsts step of this goal, we focus on effective tagging for manga character retrieval because characters are most important elements in manga. We employ the measurement “similarity” to measure the difference between line drawings representing characters. Similarity becomes higher when the two line drawings represent a same character. Similarity is obtained as weighted distance of CMRHOG features, where the weights are obtained by AdaBoost learning. We use pool-based active learning for selecting training data efficiently because the training data should be labeled by handmade which is high cost. Experimental results demonstrates the effectiveness of similarity and the selection of distinctive training samples based on pool-based active learning.