Deep learning for Fonts

Deep learning for Fonts


2016. It was a sunny morning in London, and Nischal Harohalli Padmanabha and Raghotham Sripadraj were speaking about deep learning at PyData London. Of all the things they saw that day, there was this one thing they were super curious to know: what is the font on the boards of London’s buses?

It was one of the most legible, crispest, and cleanest fonts they had ever seen. They quickly googled to figure out it was Johnston. That was easy. However, they soon began to notices a lot of instances where couldn’t determine the font so easily. Then, the big idea. Given that they are deep learning practitioners, why not use it to classify fonts? Why not have an app where you can point at an image of text and it tells you the font? Hold on! Maybe someone had already thought about this. A quick search shows two good options: Font Squirrel and What the Font. But it turns out that the results are not as expected with those apps, so Nischal and Raghotham decided to build something better: Fontastic.

Nischal and Raghotham share their experience of building a deep learning classifier with as little data as possible.

Topics include:

  • Data acquisition techniques
  • Data augmentation techniques used to generate more data and increase generalization of models
  • Determining how two or more fonts are similar
  • Which frameworks and libraries can get you the best baseline results
  • How to expose the classifier as an API
  • Date
    London, United Kingdom
    Nischal Harohalli Padmanabha
    Software Engineer by profession, filter kapi drinker by choice.

    My research interests include deep learning, large scale engineering and social interactions.