MiniDeepFont: Fonts classification CNN
Based on DeepFont Paper by Adobe
This is my final project in computer vision course #22982 of the Open University of Israel
The input file was generated from SynthText. The objective is to detect one out of three possible fonts. The network I created is based on the one Adobe suggests in their paper, minus some layers and the unsupervised learning stage, and it achieved ~95% accuracy.
When given .h5 file, the scripts create a cached fonts database, ready for training/predication. The names of these files should be set by the user.
Training
In CreateModel.py:
- Line 22: Set the list of training datasets
- Line 23: Set font cache database
- Line 24: Set the validation dataset file names (can be left empty for no validation)
- Line 25: Set the validation results filename
- Line 32: Set the model file name (.h5 suffix will be added)
Execution
python3 CreatModel.py
Testing
Execution
python3 TestModel.py
- model file name: The model .h5 file
- test set h5 file name: Test set database .h5 file
- csv result file name: Test results file
- intermediate temp file: Fonts cache database
Required packages
- matplotlib
- skimage
- tensorflow
- keras
- h5py
- numpy
- csv