Mask RCNN for Text Detection in Comics

Abstract

Object detection finds its application in a myriad of realms, from self-driving cars to face detection for security systems. One such implementation is acquiring descriptive metadata for the electronic versions of comic books hosted on the web, in order to facilitate efficient searching and cataloguing. This paper explores Mask RCNN, and its performance with respect to the detection of object instances in hand-drawn images. The scope of investigation has been restricted to object classes commonly encountered in Manga, which are, panels, text, character faces, and character bodies. The accuracy for text detection was found to be the highest.