Iterative Inference Technique for Image-to-image Translation of Tile Art Image Generation

Abstract

Tile art made by assembling small pieces of tiles is one of the artistic techniques. Tile art image generation is generating a tile art image that resembles a given digital photos and illustrations on the computer. In our previous work, we proposed a tile art image generation method using conditional generative adversarial networks. This method can generate a tile art image only by inference computation of the trained network. However, in generated tile art images, some tiles have noises and lack of edges. The main contribution of this work is to show a quality improvement technique of the generated tile art image using iterative inference. In this technique, the generated tile art image is given to the trained generator again as an input image, and this iteration of inference is repeated several times. As a result, the generated tile art images obtained by iterative inference technique have less noise and clearer edges of tiles.