Analysis of Material Representation of Manga Line Drawings using Convolutional Neural Networks

摘要

Visual perception of materials that make up objects has been gaining increasing interest. Most previous studies on visual material-category perception have used stimuli with rich information, e.g., color, shape, and texture. This article analyzes the image features of the material representations in Japanese "mange" comics, which are composed of line drawings and are typically printed in black and white. In this study, the authors first constructed a manga-material database by collecting 799 material images that gave consistent material impressions to observers. The manga-material data from the database were used to fully train "CaffeNet," a convolutional neural network (CNN). Then, the authors visualized training-image patches corresponding to the top-n activations for filters in each convolution layer. From the filter visualization, they found that the filters reacted gradually to complicated features, moving from the input layer to the output layer. Some filters were constructed to represent specific features unique to manga comics. Furthermore, materials in natural photographic images were classified using the constructed CNN, and a modest classification accuracy of 63% was obtained. This result suggests that material-perception features for natural images remain in the manga line-drawing representations. (C) 2017 Society for Imaging Science and Technology.