Graph matching based anime colorization with multiple references

Abstract

We propose a graph matching-based anime-colorization method from line drawings using multiple reference images. A graph structure of each frame in an input line drawing sequence helps to find correspondences of regions to be colorized between frames. However, it is difficult to find precise correspondences of whole frames only from a single reference image, because the graph structure tends to change drastically during the sequence. Therefore, our method first finds an optimal image from multiple reference images according to a cost function that represents shape similarity between nodes and compatibility of node pairs. While it is necessary to prepare several manually colored reference images, our method is still effective in reducing the effort required for colorization in anime production. We demonstrate the effectiveness of our method using actual images from our production.