Estimating Comic Content Using Deep Learning

Abstract

Extracting information from comic images is more difficult than that from natural images such as photographs. This is because they are expressed in black and white line drawings and the contents are often drawn using special techniques such as exaggeration and simplification. Due to these characteristics, sufficient extraction accuracy cannot be obtained using general functions of conventional image processing such as SIFT (Scale-Invariant Feature Transfer) and HOG (Histograms of Oriented Grandients).

In recent years, however, techniques of applying deep learning to comics have been proposed, contributing to a significant improvement of accuracy.

In addition, this technology can be used interpreting and estimating comic stories. This paper introduces research trends in deep learning applied to comic image processing.