Automatic Color Palette Creation from Words

We present an automatic framework to extract color palettes from words. This is a novel approach in comparison to existing solutions, e.g. manual creation or extraction from images. The associations between words and colors are deduced from a large database of 6 million tagged images using a scalable data-mining technique. The palette creation can be constrained by the user to achieve a desired hue template. We first focus on single words and then extend to entire texts. We compare our results against Adobe Kuler, a widely used online platform of manually created color palettes. We show that our approach performs slightly better than its non-automatic counterpart in terms of user’s preference rankings. This is a good result because our method is fully automatic whereas Kuler relies on users’ palettes that are manually created and annotated. Introduction Color palettes are widely used by artists and designers to colorize documents, such as posters or webpages. The colors in palettes often compose a theme with an emotion that the artists desire to evoke in the observer. Traditionally, palettes are created manually or extracted from an input image. This paper proposes a novel technique to automatically create color palettes from words. This offers to artists the opportunity to freely verbalize the relevant semantic context and instantly receive a matching color palette. Our method can handle a single or multiple words, as well as any type of words, such as objects, locations or emotions. We start with a large vocabulary of 100,000 frequently used words that we determine from Google n-grams, a byproduct of Google’s book scanning project [1]. We then download for each word 60 related images using Google Image Search and store them in a database. The images and related words are used to learn associatins between words and colors using a large scale statistical method [2, 3]. Given a specific word, the method employs a statistical significance test to assess whether its related images contain significantly more (or less) pixels of a given color. The link between a word and a color is expressed in the form of a normalized z value. We precompute these z values for all words and all colors in a color histogram. The color palette extraction takes as input one or multiple words and looks up the respective significance distributions from a database. The palette’s colors are then chosen so that they best represent the dominant regions of the words’ significance distributions. We can add optimization constraints such as a hue template to assure a palette’s color harmony. We also demonstrate that our framework is able to extract color palettes from entire texts such as Wikipedia articles. This frees artists from extracting relevant keywords if they have, e.g., a project description in the form of a paragraph. Instead, they Figure 1. The four hue templates used for this research project. We do not use more complicated templates as they have been found to be less