An application to classify handwritten Bengali graphemes.

Image source

Interact with the streamlit application here.

App preview:

This short YouTube video and this blog post describes the project.

• Dataset
• Model
• Project Structure
• Setup and Installation
• Usage
• Contributing

The dataset is introduced in this paper: A Large Multi-Target Dataset of Common Bengali Handwritten Graphemes.

The dataset is open-sourced as a part of a public Handwritten Grapheme Classification Challenge on Kaggle to benchmark vision algorithms for multi-target grapheme classification.

Link to Kaggle page: Bengali.AI Handwritten Grapheme Classification

You can read the paper summary here.

A simple Convolutional Neural Network model has been used. The training code is mostly based on this Kaggle kernel with minor tweaks.

│   .gitignore
│   app.py
│   packages.txt
│   README.md
│   requirements-dev.txt
│   requirements.txt
│
├───assets
│       app-preview.gif
│       grapheme-components.png
│       header.png
│
├───docs
│       Grapheme-paper-review.docx
│       Grapheme-paper-review.pdf
│
├───input
│   │   class_map.csv
│   │
│   ├───bengaliai-cv19
│   │       class_map.csv
│   │       class_map_corrected.csv
│   │       sample_submission.csv
│   │       test.csv
│   │       test_image_data_0.parquet
│   │       test_image_data_1.parquet
│   │       test_image_data_2.parquet
│   │       test_image_data_3.parquet
│   │       train.csv
│   │       train_image_data_0.parquet
│   │       train_image_data_1.parquet
│   │       train_image_data_2.parquet
│   │       train_image_data_3.parquet
│   │       train_multi_diacritics.csv
│   │
│   ├───kalpurush-fonts
│   │       kalpurush-2.ttf
│   ├───sample-images
│   └───upload-images
├───model
│   │   model.png
│   │
│   └───my_model
│       │   saved_model.pb
│       │
│       └───variables
└───notebooks
    └───Bengali Graphemes Multi Output CNN.ipynb

app.py - The streamlit app code, including input image preparation and inference logic.

packages.txt - List of OS level dependencies. Required for deploying app via share.streamlit.io.

requirements-dev.txt - List of required Python packages for dev environment (for training model using Jupyter and running streamlit app locally).

requirements.txt - List of Python packages required for deploying app via share.streamlit.io.

docs - Folder containing paper summary document.

input/class_map.csv - Maps the class labels to the actual Bengali grapheme components.

input/bengaliai-cv19 (gitignored) - Data files obtained from Kaggle. See Setup and Installation.

input/sample-images - Sample images used by the streamlit app when input mode is set to "Use provided images".

input/upload-images - Some samples of my own handwriting that you can upload when input mode is set to "Upload own character images".

model - Folder containing saved model and an image of the model architecture.

notebooks - Folder containing the Jupyter Notebook for model training.

Prerequisite: Python 3.7

Install required packages in a Python virtual environment:

$ pip install -r requirements-dev.txt

Next, to obtain the data from Kaggle, set up official Kaggle API from here.

After setting up the official Kaggle API, open terminal and run:

$ kaggle competitions download -c bengaliai-cv19

Unzip the downloaded file and arrange the data in the input folder as shown in the Project Structure section.

To explore the dataset or train the model, run the Jupyter Notebook - notebooks/Bengali Graphemes Multi Output CNN.ipynb

To run the streamlit app using the provided pre-trained model, run:

$ streamlit run app.py

Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.