Highly accurate mural sketch generation for Dunhuang murals with enhanced patch-GAN

Dunhuang murals are historical treasures, which have been eroded by various deteriorations during thousands of years. As a form of conservation, mural sketch drawing is widely used. However, the drawing is done manually by highly skilled experts, requiring high level of artistic appreciation and is time-consuming. A variety of algorithms for automatic mural sketch extraction were proposed. These algorithms either fail to filter the mural deteriorations effectively or barely obtain bold mural lines which are not enough for practical use. Motivated by these observations, we propose a novel deep generative model with edge guidance, stronger backbone, scaled cross-entropy and stronger supervision from multi-scale PatchGAN. With these simple yet effective enhancements, our proposed model can not only filter mural deterioration, but also generate thin lines, making this model a practical high-accuracy one. Experiments on the Dunhuang mural dataset and generalization study on non-Dunhuang murals demonstrate the effectiveness of our method.