Manga Face Detection on Various Drawing Styles Using Region Proposals-Based CNN

Abstract

Faces of characters in comic books can be used as meta-features for manga analytics. Manga character faces are not easy for a machine to detect when compared to human faces due to the high variation of drawing styles from various distinct authors. There exist several convolutional neural network-based (CNN-based) frameworks that can achieve high accuracy in an object detection task. However, their drawback is time and resource consuming to perform data modeling due to the nature of deep learning. Thus, this paper is to propose a method to develop a model using Mask R-CNN, which is one of the CNN-based frameworks, with the transfer learning technique in order to reduce training time and resources while maintaining high performance in the manga character face detection task. The proposed method could achieve the average precision of 87% in the manga character face detection tasks on both seen and unseen drawing styles. It significantly outperforms the existing conventional methods. Moreover, pre-trained weights from MS COCO dataset are transferable to manga character face detection tasks. Therefore, a well-performed manga character face detector could be developed using a limited amount of training data and time.