Data-driven Methods for Capturing and Replicating Artistic Styles in Graphical Design

捕捉和複製平面設計中藝術⾵格的數據驅動方法

Student thesis: Doctoral Thesis

View graph of relations

Author(s)

  • Quoc Huy PHAN

Related Research Unit(s)

Detail(s)

Awarding Institution
Supervisors/Advisors
Award date3 May 2016

Abstract

Artistic style has been studied by many scholars and from very different perspectives. It expands our knowledge about artists and their best practices, and helps apprentices improve their skills by learning through observation, one of the most effective methods of learning. Designers learn from the seniors by observing their designs, and musician trainees learn by listening to musical performances. In the age of the Internet, it is easier than ever for an enthusiast to find these pieces of information, which in the following, we will refer to as data. In fact, data has become the main source of experiences and practices one can rely on in this emerging information age. It is, however, not easy for an inexperienced user to find and extract useful information from the data, without having a certain degree of technological knowledge. This is one of the reasons for the thriving of artificial intelligence (AI) and machine learning (ML) in the recent years. With the help of AI, data is automatically processed, analyzed and finally transformed into an easy-to-understand format, which can be digested by any person.
The objective of this dissertation is to apply AI/ML techniques to study artistic styles by means of graphical design data. In other words, we aim to extract precious knowledge and practices from the data and use it to improve productivity in daily design tasks. Specifically, we study three cases of design data, namely, fonts, paintings and decorative objects. We develop practical applications from the studies and thoroughly evaluate their usefulness with both quantitative and qualitative experiments. For the case of font, we introduce an example-based font synthesis system, which is capable of generating complete typefaces from a few input glyphs by learning the transferring rules from font collections. For the case of paintings, we develop a novel photo-style exploration method, which was built upon robust palette data sorting and palette interpolation algorithms. Finally, we elaborate a new statistical model that well captures the styles of design in decorative object datasets and use it to recommend compatible design elements as well as reasonable compositions to novice designers.