Simplifying Sketches with Conditional GAN

In the process of creating paintings by artists, the first and foremost step is to outline the overall structures and lines of the object on paper, and then convert it into simplified sketches by scanning or redrawing. Traditional CNN-based sketch simplification models often ignore the overall semantic structure of the sketch, and have limited simplification effect on sketch with shadows and a large number of messy lines. In this paper, we present a novel sketch simplification model based on conditional GAN which learned to extract the semantic features of important lines in the original sketch and generate clear simplified sketches. We achieved this by introducing self-attention mechanism and perceptual loss. The results of multiple groups of user researches show that the simplified sketch generated by our model is superior to state-of-the art models in all aspects such as image aesthetics, line simplicity and content consistency.