This paper describes a system for detecting inappropriate images in manga books intended for reading by boys and girls. However, since a page of a manga book has multiple frames, each of which contains images of different scenes, it is necessary to determine inappropriate images for each frame of the page. In this study, we developed an inappropriate image detection system that combines a CNN for automatically extracting frames on a page as polygons and a CNN for recognizing the presence or absence of inappropriate objects in each extracted frame, specializing in detecting exposed female breasts, the most frequently appearing inappropriate image in manga books. This system has improved the efficiency and quality of inappropriate image detection, which was previously possible only by human visual inspection.
抄録全体を表示