This project uses the EfficientNetB0 CNN architecture for image classification of the American Sign Language (ASL) alphabet and 0-9 digits, with real-time detection support.

Install all required libraries from the requirements.txt file:

pip install -r requirements.txt

Download and extract the ASL dataset using the following command:

gdown https://drive.google.com/uc\?id\=1b0-MLad_AcVvbocCk7RUB2XH5Xbr7L3x

sudo apt install rar
unrar asl.rar

Before training, set up the COMET_API_KEY in a .env file inside the neuralnet directory to log metrics.

To train the model, run:

python3 ASL-Alphabet-Detection/neuralnet/train.py

Hyperparameter configurations are available in train.py.

To run the real-time detection demo:

python3 detect.py

Note: If you don't have a webcam, you can use the DroidCam app to turn your mobile phone into a webcam. Logs will be saved in the action_handler.log file.

You can use the pre-trained model best_model.pth located in the assets/ directory to perform inference.

Loss Curves	Accuracy Curves

The best model was selected based on the highest test accuracy and was trained for 25 epochs, with the best results at epoch 22.

Confusion Matrix

Feel free to report any issues you encounter.
Don't forget to ⭐ the repo!