Kannada-Handwriting-Calculator

Contour boxes (green), predicted classes (blue) and prediction probabilities (red)

A calculator that uses handwritten ಕನ್ನಡ (Kannada) digits and operators to calculate the result, using contour detection and ConvNet (Convolutional Neural Network) model prediction.

• PyTorch is used to create, train and load the state of the neural network model used for predictions.
• OpenCV and Pillow (PIL) are used to read input from the GUI canvas and to obtain contours for individual digits/operators.
• CustomTKinter is used to provide the GUI.
• The individual digits/operators are detected and their most probable target classes are predicted. The predictions are combined into a string and evaluated to get the result.
• src/ConvNet.ipynb consists of processing the data, creating the CNN model architecture and then training the model.
• src/Main.ipynb consists of loading the trained model state and using it to make predictions for the calculator app.

Requirements -

For Linux

conda create --name <env> --file requirements.txt
conda activate <env>
pip install customtkinter

For Linux / others

conda create -c conda-forge -c pytorch --name <env> --file reqs.txt
conda activate <env>
pip install customtkinter

• python 3.9.18
• pytorch 2.3.1
• opencv 4.10.0
• numpy 2.0.0
• pillow 10.3.0
• customtkinter 5.2.2
• pandas 2.2.2 [ Training only ]
• torchvision 0.18.1 [ Training only ]

Instructions -

• Clone the repo and run the Jupyter Notebook, src/Main.ipynb, or run src/Main.py.

• You can use Kannada digits ೦ ೧ ೨ ೩ ೪ ೫ ೬ ೭ ೮ ೯, operators + - × /, decimal point . and parentheses ().

• You can also use ×× for exponentiation and // for floor division.

• To train the model yourself, download the data, unzip, move it into the datasets directory, and then run the Jupyter Notebook src/ConvNet.ipynb.

  NOTE: Depending on your screen resolution, you may need to change canvas size, brush delta/displacement and brush thickness to get more accurate results.

Data

• Kannada MNIST dataset
• Symbol dateset by Irfan Chahyadi