DeepFont Paper is a technique created by Adobe.Inc to detect font from images using deep learning . They published their work as a paper for the public . Inspiring their work , I converted their thesis to a working code .
- Its trained on AdobeVFR Dataset which contains 2383 Font Categories
- Its Domain adapted CNN
- Its Learning is based upon Model Compression
The work is splited into 4 steps :
-
Dataset: Since AdobeVFR Dataset datalink is huge in size and contains lot of font categories . We created custom dataset based upon required font patches using TextRecognitionDataGenerator github. The sample folder will be available in this repo.
-
Preprocessing of Dataset: Fonts are not like objects , to have to huge spatial information to classify their features . To identify very minute feature change deepfont used certain preprocessing techniques they are
- Noise
- Blur
- Perpective Rotation
- Shading (Gradient Illumination )
- Variable Character Spacing
- Variable Aspect Ratio
-
CNN Architecture: Unlike other image classification CNN network , they followed a new schema like two subnetworks,
- Low Level Sub-Network : Learned from the composite set of synthetic and real-world data.
- High Level Sub-Network : Learns a deep classifier from the low level features For more details and clarification have a read of their paper
-
Framework ( Keras ): As its prototyping , I used Keras to build the entire pipeline . Feel free to prototype in other frameworks.
Thanks to DeepFont Team for their amazing work