🖐 ASL Sign Language Recognition

A deep learning project that recognizes American Sign Language (ASL) hand signs using a Convolutional Neural Network (CNN) built with TensorFlow/Keras.
The project includes:

• Dataset preprocessing & augmentation
• Model training with class balancing
• Image upload-based prediction using Tkinter GUI

---

📌 Features

✅ Data preprocessing with augmentation (rotation, zoom, shift)
✅ CNN model trained with class balancing
✅ Save and load trained model & labels for inference
✅ Tkinter GUI for uploading images and getting predictions
✅ Shows predicted ASL letter with confidence score

---

📂 Project Structure

ASL_Project/
│── dataset/ # ASL dataset (training images)
│── asl_train.py # Data preprocessing + model training script
│── asl_gui.py # Tkinter GUI for predictions
│── best_asl_model.h5 # Trained CNN model
│── class_labels.pkl # Saved class labels
│── README.md # Project documentation

---

📊 Dataset

We use the ASL Alphabet Dataset, which contains folders for each ASL sign (letters A–Z, plus additional classes like "space", "nothing", etc.).

Dataset Structure:

dataset/ │── A/ │── B/ │── C/ │── ...

• Training Images: High-resolution images of hands forming each ASL sign.
• Class Count: Matches the number of folders in the dataset.
• Important Note:
  • Do NOT apply horizontal flipping for ASL gestures — flipping can change the meaning of the sign.

---

🔄 Workflow

1️⃣ Data Preprocessing

• Rescale images to 1./255 for normalization.
• Resize all images to 64×64 pixels for consistent input shape.
• Apply data augmentation:
  • Rotation (±20°)
  • Zoom (up to 20%)
  • Width/Height shift (up to 10%)
  • Shearing (15%)
• Split dataset into:
  • 80% Training
  • 20% Validation

2️⃣ Class Label Preparation

• Extract all folder names as class labels.
• Save labels into class_labels.pkl for use during prediction.

3️⃣ Model Training

• CNN architecture:
  • 3 convolutional layers with Batch Normalization
  • MaxPooling after each conv block
  • Dropout for regularization
  • Fully connected Dense layer before output
• Loss Function: Categorical Crossentropy
• Optimizer: Adam
• Metrics: Accuracy
• Training Strategy:
  • Early Stopping (patience=2)
  • Reduce Learning Rate on Plateau
  • Save best model as best_asl_model.h5

4️⃣ Prediction Phase

• Load best_asl_model.h5 and class_labels.pkl.
• Accept uploaded image from GUI.
• Preprocess image:
  • Resize to 64×64
  • Normalize pixels (1./255)
• Model outputs:
  • Predicted Class
  • Confidence Score