Cartoon Character Recognition Using Concentric Multi-Region Histograms of Oriented Gradients

Abstract

Comic books are a kind of serial narrative artwork made up of comic pages. As an essential part of comics, cartoon characters appear throughout the whole series. Therefore, the recognition of cartoon characters is useful for many applications of comics. Normally, images of the same character are similar but with different representations in different scenes, such as facial expressions, poses, and viewpoints, which make them difficult to be recognized. In contrast to human being, besides face regions, there are many other parts offering the identification features for cartoon characters. In this paper, we focus on cartoon character recognition and propose Concentric Multi-Region model to explore the significant features from the parts around face regions. Histograms of Oriented Gradients (HOG) is utilized for the description of regions, and the AdaBoost algorithm is applied to obtain a new descriptor named Concentric Multi-Region Histograms of Oriented Gradients (CMR-HOG). In the experiments, 17 labeled cartoon characters are applied. Compared to other face and object recognition methods only based on face regions, the proposed method shows better performance. In addition, we proved its scalability for cartoon character recognition.