Typeface Semantic Attribute Prediction from Rasterized Font Representations

Typefaces comprise an important component of the aesthetic quality of any written work. With the rise of printers and digital word processors has come a tremendous diversity of typefaces, each slightly different in visual effect and design. With this growing number of choices, there has been an increasing need to intelligently categorize and characterize typefaces for searching, grouping, pairing, and more. Most approaches thus far have categorized typefaces by their font styles—for example, ‘bold’ or ‘italic’—as well as other typographic attributes, such as ‘all-capitals’ or ‘cursive’. However, while these characteristics can be descriptive for some fonts, these aspects fail to capture some of the more aesthetic qualities—what we call semantic attributes—that we as humans associate with typefaces in abstract senses, such as ‘artistic’, ‘boring’, or ‘gentle’, which may be desirable when one is looking for typefaces that have particular semantic attributes. We are thus interested in applying machine learning to build on existing work for generalizing the representation and characterization of typeface semantic attributes beyond existing semantic attribute datasets.