Glyph Cropping for Font Generation
Objective: To develop an automated glyph isolation model to efficiently crop and extract glyphs from Pecha images, facilitating the creation of fonts from various Pecha publications.
Background: Previously, glyph isolation was performed manually by annotating the bounding polygons of glyphs. While this method produced accurate results, it was not scalable for larger datasets. To overcome this challenge, we are transitioning to a machine-learning approach to automate glyph cropping.
Current Status:
- Completed Work: We have successfully created a font from the Pecha glyphs of the Derge publication.
- Ongoing Work: We are working with incomplete glyph data from four other Pecha publications.
Solution Approach: To develop a custom glyph cropping model that will be trained on our existing dataset of cropped glyph images. This model aims to automate the extraction of glyphs from Pecha images, improving scalability and efficiency in font generation.
Model Development and Testing
Model Overview: We have successfully developed and tested a custom glyph cropping model designed to automate the isolation of glyphs from Pecha images. This model is a key component in generating fonts from various Pecha publications.
Development:
- Model Architecture: The model was built using a U-Net Convolutional Neural Network (CNN) with standard layers.
- Training: The model was trained on a dataset of cropped glyph images as the target image and a pair of mixed glyph images and conditional images as the input images.
Testing and Results:
-
Performance: The model has been tested on both the training dataset and a separate validation set. It consistently yielded accurate and reliable results.
-
Observations: The model demonstrated effective glyph isolation with high precision, successfully extracting glyphs from Pecha images.
-
Integration: Implement the model in the production pipeline for automated glyph extraction.
-
Dataset creation and training
