This project is not an OCR. Instead, it contains a dataset of perfect test images of Sinhala characters that were generated, along with a test Convolutional Neural Network (CNN) based on LeNet. The purpose is to train the dataset to identify images that contain Sinhala characters. Caffe was used to implement and train the network. Feel free to use the provided dataset for your own purposes.
.
├── commands.sh # contains useful Caffe commands
├── imgs # dataset
│ ├── adjusted # training images with label files (TRAIN & TEST phase)
│ ├── labels_unicode.txt
│ └── unicode # a new dataset to evaluate the trained model
├── lenet.prototxt # TEST model
├── lenet_solver.prototxt # Caffe solver
├── lenet_train_test.prototxt # TRAIN model
├── map.txt # mapping of labels / Unicode char / other Sinhala fonts
├── model # output models / solverstate directory
├── node_modules
├── README.md
├── test.py # Python script to evaluate the trained model
└── textgen.js # script to generate Sinhala characters
Once you have defined lenet_train_test.prototxt and configured solver lenet_solver.prototxt, start training using the following commands. You can view commands.sh file for useful Caffe commands.
Use test.py to test the trained model on your new images.
To generate Sinhala characters, you need to install the following dependencies:
sudo apt-get install libcairo2-dev libjpeg8-dev libpango1.0-dev libgif-dev build-essential g++
Make sure all the fonts mentioned in the code are installed on your system and that font names don't have any spaces or escape characters. To generate Sinhala characters, run the following commands:
npm install
node textgen.js
The created image will be saved in imgs/ directory.