Capturing Fonts in the Wild

Abstract

Editing text in photographs requires the ability to find the same font, which is impossible in many settings, such as historical or manually painted text. We present a method of extracting the font from a single photographed word, without relying on the retrieval of similar fonts. A deep net extracts style information and constructs the font for all characters, enabling novel applications in image editing, font creation, and the addition of language-specific characters with diacritics to existing fonts. A qualitative user study shows that this method improves convincing font capture by over 500% over prior work.