Universal Font Recognition (DeepFont)

This is a PytTorch implementation of the method proposed in DeepFont to recognize the font from text images using deep learning. This ropo can be applied for any language and font.

An automatic font image generator is also provided in this repo using TextRecognitionDataGenerator. You may refer to it if you want to customize the generation mechanism.

Installation

Clone the Code

git clone https://github.com/twelfth-star/universal-font-recognition.git
cd universal-font-recognition

Create the Environment

conda create -n font_recog python=3.8.13
conda activate font_recog

Install Required Packages

pip install -r requirements.txt

Note: If you want to use GPU, you should install PyTorch consistent with your CUDA version (use command nvidia-smi to check your CUDA version). Detailed instructions may be found in its official website.

How to Train Your Own Model

1. Download the font files (usually .ttf, .ttc, otf, etc.) and put them into ./dataset/fonts (or anywhere else you like).

2. (Optional) Put the real images with text, whose font is included in the font files you prepared in Step 1, into ./dataset/real_images (or anywhere else you like). No specific labels are needed.

3. Edit the settings in .font_recognition/main.py based on your preference.

4. (Optional) Edit the image generation mechanism in function generate_batch_images in general_code/image_generation.py based on your preference. You may refer to the official wiki of TextRecognitionDataGenerator

5. Enter ./font_recognition using cd ./font_recognition. Run the code using python ./main.py. The images for training will be automatically generated and the model will be trained.

6. By default, the model will be saved as ./dataset/models/CNN_{language}_{num_fonts}.pth. A json file indicating the map from font name to font ID will also be saved as ./dataset/models/font_dict_{language}_{num_fonts}.json

How to Use Trained Model

Brief Introduction of DeepFont

Recognition of Other Text Attributes