Abstract

As font is one of the core design concepts, automatic font identification andsimilar font suggestion from an image or photo has been on the wish list ofmany designers. We study the Visual Font Recognition (VFR) problem, and advancethe state-of-the-art remarkably by developing the DeepFont system. First ofall, we build up the first available large-scale VFR dataset, named AdobeVFR,consisting of both labeled synthetic data and partially labeled real-worlddata. Next, to combat the domain mismatch between available training andtesting data, we introduce a Convolutional Neural Network (CNN) decompositionapproach, using a domain adaptation technique based on a Stacked ConvolutionalAuto-Encoder (SCAE) that exploits a large corpus of unlabeled real-world textimages combined with synthetic data preprocessed in a specific way. Moreover,we study a novel learning-based model compression approach, in order to reducethe DeepFont model size without sacrificing its performance. The DeepFontsystem achieves an accuracy of higher than 80% (top-5) on our collecteddataset, and also produces a good font similarity measure for font selectionand suggestion. We also achieve around 6 times compression of the model withoutany visible loss of recognition accuracy.