🎭 Emotion Detector with CNN

This project implements a Convolutional Neural Network (CNN) for real-time facial emotion recognition using the FER2013 dataset. The model is capable of detecting 7 emotions from facial expressions: angry, disgust, fear, happy, sad, surprise, and neutral.

---

📌 Project Structure

• training.py: Script to train the CNN model on the FER2013 dataset.
• model.py: Definition of the neural network architecture using PyTorch.
• app.py: Real-time emotion recognition using OpenCV and a webcam.
• model.pth: Trained model weights (exported after training).

---

🧠 Model Architecture

The architecture consists of 4 convolutional blocks (Conv2D + BatchNorm + ReLU + MaxPooling), followed by a fully connected head:

self.conv1 → self.conv2 → self.conv3 → self.conv4 → flatten → fc1 → fc2

Dropout is used to reduce overfitting. The final layer outputs a 7-class prediction using softmax.

---

📊 Training

• Dataset: FER2013 (48x48 grayscale facial images)
• Accuracy achieved: ~56% on validation
• Loss function: CrossEntropyLoss
• Optimizer: Adam, learning rate = 0.001
• Epochs: 50
• Data Augmentation:
  • Random horizontal flip
  • Random rotation

Training was performed in Google Colab using GPU acceleration. The final model was exported to a .pth file and used locally for inference.

---

🔍 Inference Pipeline

1. Open webcam feed using OpenCV.
2. Detect face using Haar Cascades.
3. Preprocess the detected face:
   • Convert to grayscale
   • Resize to 48x48
   • Convert to PyTorch tensor
4. Pass the image through the trained model.
5. Display the predicted emotion on screen using cv2.putText.

---

⚙️ Requirements

Install dependencies with:

pip install -r requirements.txt

Main libraries used:

• torch
• torchvision
• opencv-python
• imutils

---

🚀 How to Run

• To train the model:

python training.py

• To run the real-time emotion detector:

python app.py

Make sure model.pth is in the same directory as app.py.

---

📝 Notes

• The model was trained on grayscale images resized to 48x48 pixels, so input preprocessing must match.
• You can further improve performance by:
  • Using a deeper CNN (e.g., ResNet)
  • Applying more aggressive data augmentation
  • Training on a larger or more balanced dataset
• Emotion detection is approximate and works best under good lighting conditions.

---

📸 Preview

---

📄 License

This project is licensed under the MIT License – see the LICENSE file for details.