DeepFont: Identify Your Font from An Image Zhangyang (Atlas) Wang Joint work with Jianchao Yang, Hailin Jin, Jon Brandt, Eli Shechtman, Aseem Argawala, and Thomas Huang

DeepFont: Product & Press

Problem Definition Seen a font in use and want to identify what it is?

Problem Definition Font recognition: recognize font style (typeface, slop, weight, etc) automatically from real-world photos Why it matters? Highly desirable feature for designers Design library collection Design inspiration Text editing

Challenges An extremely large-scale recognition problem Over 100,000 fonts claimed on myfonts.com in their collection Beyond object recognition: recognizing subtle design styles. Extremely difficult to collect labeled real-world training data Has to rely on synthetic training data BIG mismatch between synthetic training and real-world testing

Solution Deep convolutional neural network? Effective at large-scale recognition Effective at fine-grained recognition Data-driven Problem: huge mismatch between synthetic training and real-world testing Task-Specific Data augmentation Decomposition-based deep CNN for domain adaptation

The AdobeVFR Dataset Synthetic training set 2383 fonts from Adobe Type Library (extended to 4052 classes later) 1000 synthetic English word images per font ~2.4M training images Real-world testing set 4383 labeled images, covering 671 fonts out of 2383 Over 100K unlabeled images …………………………………………… The first large-scale benchmark for the task Both synthetic and real-world text images Also good for fine-grain classification, domain adaption, understand design styles

Deep Convolutional Neural Network Following the benchmark structure?

Domain Mismatch Direct training on synthetic data and testing on real-world data (Top-5 accuracy) Need domain adaptation to minimize the gap between synthetic training and real-world testing! Synthetic Real-World Training 99.16% NA Testing 98.97% 49.24%

Data Augmentation Common degradations Noise, blur, warping, shading, compression artifacts, etc

Data Augmentation Common degradations Noise, blur, warping, shading, compression artifacts, etc Special degradations Aspect ratio squeezing: squeeze the image using a random ratio in [1.5, 3.5] in the horizontal direction.

Data Augmentation Common degradations Noise, blur, warping, shading, compression artifacts, etc Special degradations Aspect ratio squeezing: squeeze the image using a random ratio in [1.5, 3.5] in the horizontal direction. Random character spacing: render training text images with random character spacing

Data Augmentation Common degradations Noise, blur, warping, shading, compression artifacts, etc Special degradations Aspect ratio squeezing: squeeze the image using a random ratio in [1.5, 3.5] in the horizontal direction. Random character spacing: render training text images with random character spacing Inputs to the network: random 105x105 crops

Effects of Data Augmentation Synthetic 1-4: common degradations Synthetic 5-6: special degradations Synthetic 1-6: all degradations On the right: MMD between synthetic and real-world data responses

Beyond Data Augmentation Problems Cannot enumerate all possible degradations, e.g., background and font decorations. May introduce degradation bias in training Design the learning algorithm to be robust to domain mismatch? Mismatch already happens in the low-level features Tons of unlabeled real-world data

Network Decomposition for Domain Adaptation Unsupervised cross-domain sub-network Cu (N layers) Supervised domain-specific sub-network Cs (7-N layers)

Network Decomposition for Domain Adaptation Train sub-network Cu in a unsupervised training using stacked convolutional auto encoders, with both synthetic data and unlabeled real-world data. Fix sub-network Cu, and train sub-network Cs in a supervised way, using the labeled synthetic data.

Quantitative Evaluation 4383 real-world test images collected from font forums. Model Augmentation? Decomposition? Real-World Test (Accuracy) Top 1 Top 5 LFE Y Na 42.56% 60.31% DeepFont N 42.49% 49.24% 66.70% 79.28% 71.42% 81.79% Varying the layer number K of unsupervised network Cu K 1 2 3 4 5 Training 91.54% 90.12% 88.77% 87.46% 84.79% 82.12% Testing 79.28% 79.69% 81.79% 81.04% 77.48% 74.03%

Successful Examples

Failure Examples

Model Compression For a typical CNN, about 90% of the storage is taken up by the dense connected layers Matrix factorization methods are considered for compressing parameters in linear models, by capturing nearly low-rank property of parameter matrices. The plots of eigenvalues for the fc6 layer weight matrix in DeepFont. This densely connected layer takes up 85% of the total model size.

Model Compression During training, we add a low rank constraint on the fc 6 (rank < k) layer In practice, we adopt very aggressive compression on all fc layers, and obtained a mini-model with ~40 MB in storage, with a compression ratio >18, and (top-5) performance loss ~3%. Take-Home Points: FC layers can be highly redundant. Compressing them aggressively MIGHT work well. Joint Training-Compression performs notably better than two-stage.

In Adobe Product: Recognize Fonts from Images

In Adobe Product: Photoshop Prototype

Text Editing Inside Photoshop

Text Editing Inside Photoshop

In Adobe Product: Discover Similarity between Fonts Font inspiration, browsing, and organization

In Adobe Product: Discover Similarity between Fonts Font inspiration, browsing, and organization

For more information & dataset download: Thank you! For more information & dataset download: atlaswang.com/deepfont.html