Style transfer based artistic transformation using Albert Gulk’s pencil drawings

Abstract

Style transfer is the process of retouching images to appear in the artistic style of another image. Researchers have conducted intensive work using advanced machine learning methods and high level mathematical models in order to transfer style of artworks. One of the challenges, which still exists, is to perform style transfer which preserves the shapes of the content image as much as possible while transferring finer style patterns. In this thesis, we investigate how state-ofthe- art style transfer models can be improved such that the shape of the content image is less disturbed during the process. For this purpose we introduce a discrete wavelet transform based neural style transfer pipeline. Experiments were conducted using the artworks of Albert Gulk, a famous Estonian artist of the Kursi school. A challenge in Albert Gulk’s pencil drawings is that they are monotone works, hence using any basic style transfer will significantly disturb the shape of content image. Experimental results show that the proposed wavelet based neural style transfer approach can preserve the shape of content when monotone artworks are used as style images.