Stylisation Through Strokes

  • Sameh Hussain

Student thesis: Doctoral ThesisPhD


This dissertation presents research outputs that address the general task of learning formal elements of an artistic style and reproducing them in novel outputs.

Prior works have presented several limitations. Works in computer graphics (\eg~Non-Photorealistic Rendering (NPR) and media emulation) are highly specialised. Machine learning algorithms (e.g. Neural Style Transfer (NST) and Neural Media Emulation) require large datasets and cannot adequately reproduce the art style depicted in the examples.

The contributions in this dissertation address the problem by building machine learning algorithms around art theory. We achieve this by avoiding pixel-level optimisation and instead focusing on producing art using strokes.

Our research has resulted in four contributions, consisting of three research contributions and one engineering contribution. Our first contribution addresses the production of hatched drawings according to an artist-defined hatching style. Our second contribution targets learning the manner of stroke making, i.e. stroke style, from artist-drawn examples. Our third contribution is a set of experimental results that assess the impact of stroke style on human perception of automatically generated art. Our last research contribution focuses on learning art media from photographs. In addition to our research contributions, we will present our engineering contribution that integrates our research contributions into an interactive tool that can produce stylised outputs in a fully automated or user-guided fashion.

Our results demonstrate the value of using art theory to guide the development of machine learning algorithms that target artistic style. Each of our contributions requires only a handful of examples to learn from an artist. Furthermore, our models can extrapolate from the examples and produce novel outputs.

In conclusion, our contributions address the specificity of prior works in computer graphics (e.g. NPR and media emulation) whilst also addressing the inability of machine learning algorithms to learn artistic style adequately. Professional artists have expressed a positive outlook on our interactive tool and the impact of our contributions on industrial processes.
Date of Award24 May 2023
Original languageEnglish
Awarding Institution
  • University of Bath
SupervisorPeter Hall (Supervisor), Christian Richardt (Supervisor) & Andrew Vidler (Supervisor)

Cite this