Doodle Master: A Doodle Beautification System Based on Auto-encoding Generative Adversarial Networks

For those people without artistic talent, they can only draw rough or even awful doodles to express their ideas. We propose a doodle beautification system named Doodle Master, which can transfer a rough doodle to a plausible image and also keep the semantic concepts of the drawings. The Doodle Master applies the VAE/GAN model to decode and generate the beautified result from a constrained latent space. To achieve better performance for sketch data which is more like discrete distribution, a shared-weight method is proposed to improve the learnt features of the discriminator with the aid of the encoder. Furthermore, we design an interface for the user to draw with basic drawing tools and adjust the number of reconstruction times. The experiments show that the proposed Doodle Master system can successfully beautify the rough doodle or sketch in real-time.