Artpdgan: Creating artistic pencil drawing with key map using generative adversarial networks

Abstract

A lot of researches focus on image transfer using deep learning, especially with generative adversarial networks (GANs). However, no existing methods can produce high quality artistic pencil drawings. First, artists do not convert all the details of the photos into the drawings. Instead, artists tend to use strategies to magnify some special parts of the items and cut others down. Second, the elements in artistic drawings may not be located precisely. What’s more, the lines may not relate to the features of the items strictly. To address above challenges, we propose ArtPDGAN, a novel GAN based framework that combines an image-to-image network to generate key map. And then, we use the key map as an important part of input to generate artistic pencil drawings. The key map can show the key parts of the items to guide the generator. We use a paired and unaligned artistic drawing dataset containing high-resolution photos of items and corresponding professional artistic pencil drawings to train ArtPDGAN. Results of our experiments show that the proposed framework performs excellently against existing methods in terms of similarity to artist’s work and user evaluations.