Learning design: Aesthetic models for color, layout, and typography.

摘要

This thesis presents aestheticmodels for three key components of graphic design: color compatibility, layout creation, and font selection. We develop thesemodels using a variety of techniques, including machine learning and data mining, crowdsourcing, and optimization, and demonstrate their use in practical design applications. We .rst model color compatibility from large datasets, including analysis of existing color theories, a learned model for predicting compatibility, and several applications using this model. We then use a collaborative .ltering approach tomodel color preferences for individual users. Secondly, we present an energy-basedmodel for single-page design layouts derived from design principles, using novel analysis algorithms. We show how this model can be trained using existing layouts, and demonstrate its use on various tasks including style-based synthesis and retargeting. We then extend this model to enable an interactive, suggestion-based layout tool. Finally, wemodel fonts using attributes and similarity data from crowdsourced studies, then use these models to develop exploratory font selection interfaces.