Emotion Detection using ResNet18

Welcome to the Emotion Detection project! This project leverages the power of deep learning to recognize emotions from facial expressions using a Convolutional Neural Network (CNN) based on the ResNet18 architecture. Whether you're a data scientist, a machine learning enthusiast, or just curious about AI, this project is a great way to explore the fascinating world of emotion recognition.

Project Highlights

• Transfer Learning: Utilizes a pre-trained ResNet18 model, fine-tuned for emotion detection.
• Data Augmentation: Enhances model robustness with a variety of transformations.
• Custom Modifications: Tailored for grayscale images, with selective layer freezing for efficient learning.
• Performance Tracking: Includes detailed metrics and visualizations to monitor model performance.

Getting Started

Prerequisites

To get started, you'll need the following Python packages:

pip install torch torchvision numpy pandas opencv-python Pillow scikit-learn matplotlib seaborn tqdm

Dataset

The model is trained on the FER2013 dataset, which is structured as follows:

/kaggle/input/fer2013/
├── train/
│   ├── angry/
│   ├── disgust/
│   ├── fear/
│   ├── happy/
│   ├── neutral/
│   ├── sad/
│   └── surprise/
└── test/
    ├── angry/
    ├── disgust/
    ├── fear/
    ├── happy/
    ├── neutral/
    ├── sad/
    └── surprise/

Model Architecture

• Base Model: ResNet18 with ImageNet weights.
• Modifications:
  • First convolutional layer adjusted for single-channel (grayscale) input.
  • Final fully connected layer replaced to output seven emotion classes.
  • Selective layer freezing to retain learned features while adapting to new data.

Data Augmentation

To improve generalization, the following augmentations are applied:

• Random horizontal flips
• Random rotations up to 10 degrees
• Random affine transformations
• Random sharpness adjustments
• Random autocontrast
• Normalization to standardize input

Training the Model

The training process involves:

1. Initialization: Load and modify the ResNet18 model.
2. Optimization: Use AdamW optimizer with differential learning rates for different layers.
3. Learning Rate Scheduling: Adjust learning rates based on validation loss.
4. Gradient Clipping: Prevent exploding gradients by clipping them.
5. Early Stopping: Save the best model based on validation accuracy.

Here's a snippet to kick off training:

model = create_resnet18_model(num_classes=7).to(device)
history = train_and_validate(model, train_loader, test_loader, num_epochs=10, learning_rate=0.0001)

Visualizing Results

After training, visualize the model's performance:

• Loss and Accuracy Curves: Track training and validation loss/accuracy over epochs.
• Confusion Matrix: Understand model predictions across different emotions.
• Classification Report: Detailed precision, recall, and F1-score for each emotion.

Example Visualizations

• Training and Validation Loss: Shows how the model learns over time.
• Confusion Matrix: Provides insights into which emotions are often confused.
• Accuracy Over Time: Visualizes improvements in model accuracy.

Performance Metrics

• Best Training Accuracy: 74.52%
• Best test Accuracy: 66.88%
• Detailed Classification Report: Available in the output section of the notebook.

Future Work

• Explore other architectures like EfficientNet or Vision Transformers.
• Implement real-time emotion detection.
• Enhance data preprocessing and augmentation techniques.
• Experiment with ensemble methods for improved accuracy.

Conclusion

This project demonstrates the potential of deep learning in emotion recognition. By leveraging transfer learning and data augmentation, the model achieves decent accuracy in classifying emotions from facial expressions.