Manga character clustering with DBSCAN using fine-tuned CNN model

Manga (Japanese comic) is popular content worldwide. In Japan, e-comic accounts for about 80% of e-book market. In recent years, metadata extraction from manga image has been studied for providing e-comic service. Manga character is one of the important contents for story understanding. In conventional research, some character identification methods are proposed those classify characters' face images using k-means clustering. However, there are two problems. First, kmeans method needs to specify the number of clusters, however the number of characters in target manga images is commonly unknown. Second, manga includes characters with few appearing, so it is difficult to classify characters with high purity. To solve these problems, we propose clustering method using DBSCAN which decides number of clusters automatically and is robust to noise data. In our prior research, it is experimented that character face clustering using DBSCAN and general CNN features. However, general CNN model is difficult to capture detailed features of manga characters. In this paper, we apply DBSCAN to fine-tuned CNN with manga character faces to improve the clustering accuracy. We also compare the optimal parameter determination method of DBSCAN. Experimental results showed that the dimensional reduction using Kernel PCA and UMAP is effective. In addition, we confirmed the validity of proposed method that determining the parameters of DBSCAN based on the slope changing of k-distance graph.