Nepali Handwritten Character Classification Using CNN

Project Description:

This project aims to develop a deep learning model to classify handwritten Nepali characters into 46 different classes using Convolutional Neural Networks (CNNs). The dataset contains 92,000 grayscale images, each representing one of the Nepali characters (क, ख, ग, ... + १, २, ...). The goal is to develop and train a deep learning model that can accurately classify these characters

• Dataset Total Images: 92,000
• Training Images: 78,200
• Test Images: 13,800
• Image Size: 32x32 pixels
• Number of Classes: 46 (includes both characters and numbers)

Approach:

Model:

The model is a CNN with three convolutional layers followed by max-pooling and fully connected layers for classification.

Training:

The model is trained using the training set and evaluated on the test set using accuracy as the primary metric.

Visualization:

The project includes visualization scripts to display model predictions on a grid of test images.

Tech Stack Used:

• PyTorch: For building, training, and evaluating the CNN model.
• NumPy: For loading and manipulating the dataset.
• Matplotlib: For visualizing the predictions.
• torchvision: For dataset transformation.

File Structure

• │── main.ipynb # Main Jupyter Notebook containing the whole code
• │── dataset/ # Directory containing the dataset
• │── requirements.txt # List of dependencies required for the project
• │── README.md # Project documentation (this file)

Execution

Installing Dependencies

Make sure you have Python installed. Then, install the required dependencies using the requirements.txt file :

pip install -r requirements.txt

How To Run

• Run the main.ipynb File. This code is specifically designed to be run in a Jupyter Notebook (.ipynb format). Therefore, it is essential that you run the project using the Jupyter Notebook or any environment that can run .ipynb files.

• To run the project: Open the main.ipynb. Execute each cell step-by-step in the order they appear.

Running on Google Colab:

If you prefer running on Google Colab, don't forget to mount the google drive and change the dataset path to match your Google Drive path.

Model Training and Results

• Epochs: The model was trained across 5 epochs. An epoch represents one complete cycle through the entire training dataset, which helps the model to improve its learning.
• Training Accuracy: At the end of the 5th epoch, the model achieved a training accuracy of 97.60%.
• Test Accuracy: The final test accuracy was 98.49%, indicating the model's effectiveness on data it hasn't previously seen.
• Loss: The loss at the conclusion of the training was 0.0748 indicating that the model's predictions are closely aligned with the actual outcomes.
• Total Images Trained: Over the 5 epochs, the model processed a total of 391,000 images.

Conclusion

The model successfully classifies Nepali handwritten characters with a high accuracy of 98.49% on the test data after training for 5 epochs. The results show that the model is effective at generalizing to unseen data, making it a reliable solution for recognizing Nepali characters.

Authors:

• Anoma Tuladhar

• Nairiti Rai

• Pradeep Bhatta

• Samikshya Dahal