Facial Emotion Recognition using Convolutional Neural Networks

Project Overview

This project implements a deep learning model for facial emotion recognition using the FER2013 dataset, demonstrating advanced techniques in computer vision and machine learning.

Project Objectives

• Develop a robust Convolutional Neural Network (CNN) for classifying facial expressions
• Explore data augmentation techniques to improve model generalization
• Implement advanced deep learning strategies for image classification

Dataset Details

FER2013 Dataset

• Source: Kaggle FER2013 Dataset
• Image Specifications:
  • 48x48 pixel grayscale images
  • Faces automatically registered and centered

Emotion Categories

• 0: Angry
• 1: Fear
• 2: Happy
• 3: Neutral
• 4: Sad
• 5: Surprise

Methodology

1. Data Preprocessing and Augmentation

Why Data Augmentation?

• Reduce Overfitting: By introducing variations in training data
• Increase Training Examples: Each transformation adds new learning samples
• Improve Model Robustness: Helps model generalize to real-world variations

Augmentation Techniques:

• Rotation (±20 degrees)
• Width/Height Shifts (20%)
• Horizontal Flipping
• Zoom Range (20%)
• Rescaling to normalize pixel values

2. CNN Architecture Design

Network Structure Rationale

• Convolutional Blocks: Progressively extract spatial hierarchical features
  • Initial blocks: Extract basic features
  • Intermediate blocks: Capture complex patterns
  • Advanced blocks: Learn high-level representations

Key Architectural Choices:

• Batch Normalization: Stabilize learning
• Spatial Dropout: Prevent feature co-adaptation
• Multiple Dense Layers: Combine and reduce dimensionality

3. Training Strategy

• Optimizer: Adam (adaptive learning rate)
• Loss Function: Categorical Cross-Entropy
• Early Stopping: Prevent overfitting by monitoring validation loss

Performance Metrics

Model Performance

• Test Accuracy: 65%
• Per-Class Performance Highlights:
  • Highest Accuracy: Happy (86%)
  • Challenging Classes: Fear, Sad

Installation and Setup

# Clone the repository
git clone https://github.com/yourusername/facial-emotion-recognition.git

# Create virtual environment
python -m venv venv
source venv/bin/activate  # On Windows, use `venv\Scripts\activate`

# Install dependencies
pip install -r requirements.txt

# Run the notebook or script
jupyter notebook src/main.ipynb