FillingGAN: A Deep Learning-based Automatic Coloring Technique from Line Drawings for Developing Coloring Game Contents
GAN, Stylization, Line drawing, Coloring Game, Feature Extraction
Jeongin LEE, Heekyung YANG, Kyungha MIN
In this paper, we contribute to the field of game by presenting FillingGAN, an automatic coloring framework using a generative adversarial network (GAN). FillingGAN is devised to generate a coloring image from a line drawing by filling empty regions between the lines in the line drawing image from the coloring styles learned from sample coloring images. Our model consists of two style extracting modules and a GAN model. The style extracting modules are designed as an auto-encoder that extracts feature from input images. One module extracts color styles from coloring sample images and the other extracts structure from a line drawing. FillingGAN executes coloring process by applying the coloring styles from coloring samples to a line drawing. It determines the similarity between the generated coloring image and the input line drawing, and calculate the perceptual loss to preserve the structural similarity. FillingGAN generates coloring images with preserved details by adjusting the weights between the structural feature vectors and the style feature vectors. As a result, it generates a stylized image without distorting the unique features of a line drawing. Our framework can be improved to apply various styles including artistic media strokes to line drawings.