🎭 Real-Time Facial Emotion Recognition with ResNet50

High-performance real-time facial emotion recognition system using ResNet50 architecture, trained on FER2013 dataset.

Features • Installation • Usage • Performance • Dataset

---

🎥 Demo

Emotion_Recognition.mp4

Real-time emotion detection running at 30+ FPS

---

📋 Overview

A production-ready system that identifies 7 distinct facial emotions through webcam feeds in real-time. Built on custom ResNet50 architecture achieving 65.59% validation accuracy on the FER2013 dataset.

✨ Key Features

🚀 Real-time Processing 30+ FPS on GPU Optimized inference pipeline FP16 precision support ⚡ High Performance 65.59% validation accuracy ~80% training accuracy Minimal latency

🎯 Seven Emotions Neutral • Happiness • Sadness Surprise • Fear • Disgust • Anger 🛠️ Production Ready Temporal smoothing Threaded capture MediaPipe face detection

---

🏗️ Architecture

Model: Custom ResNet50 with Bottleneck blocks [3, 4, 6, 3]

                                            Input (224×224×3)
                                                    ↓
                                Conv2dSame(7×7) → BatchNorm → ReLU → MaxPool
                                                    ↓
                                    ResBlock Layer 1 (64 channels)
                                                    ↓
                                    ResBlock Layer 2 (128 channels)
                                                    ↓
                                    ResBlock Layer 3 (256 channels)
                                                    ↓
                                    ResBlock Layer 4 (512 channels)
                                                    ↓
                            AdaptiveAvgPool → FC(2048→512) → FC(512→7) → Softmax
                                                    ↓
                                        Output (7 emotion classes)

🎓 Training Configuration

Parameter	Value
Optimizer	Adam (lr=0.001, weight_decay=0.0001)
Scheduler	ReduceLROnPlateau (factor=0.5, patience=3)
Batch Size	64
Epochs	50
Augmentation	Flip, Rotation (±15°), Brightness/Contrast (0.8-1.2x)

---

📊 Performance

Training Progress Visualization

📈 Metrics

Metric	Value	Description
🎯 Validation Accuracy	65.59%	FER2013 test set
📚 Training Accuracy	80.00%	FER2013 train set
⚡ Speed (GPU)	30+ FPS	NVIDIA GPU (FP16)
💻 Speed (CPU)	10-15 FPS	Intel/AMD CPU
📦 Model Size	90 MB	PyTorch .pt format
🔢 Parameters	~23.5M	Trainable params

---

📚 Dataset

FER2013 - Facial Expression Recognition

Dataset Statistics: 📊 Training: ~28,709 images ✅ Testing: ~3,589 images 🎭 Classes: 7 emotions 📐 Resolution: 48×48 → 224×224

Emotion Distribution: 😐 Neutral 😊 Happiness 😢 Sadness 😲 Surprise 😨 Fear 🤢 Disgust 😠 Anger

Note: Images are grayscale 48×48 pixels, upscaled to 224×224 RGB for training.

---

🚀 Installation

Prerequisites

✓ Python 3.8+
✓ CUDA-compatible GPU (recommended)
✓ Webcam for real-time inference

Quick Setup

1️⃣ Clone repository

git clone https://github.com/yourusername/FER2013-ResNet50-Emotion-Recognition.git
cd FER2013-ResNet50-Emotion-Recognition

2️⃣ Install uv (fast package manager)

# Windows
powershell -c "irm https://astral.sh/uv/install.ps1 | iex"

# macOS/Linux
curl -LsSf https://astral.sh/uv/install.sh | sh

3️⃣ Create environment and install dependencies

uv venv
source venv/bin/activate  # Windows: venv\Scripts\activate
uv pip install -r requirements.txt

4️⃣ Add pre-trained model

Place the trained model file in the models/ directory:

models/FER_static_ResNet50_AffectNet.pt

💡 Tip: Model file size ~90MB. Available on request or train your own using the notebook.

---

💻 Usage

🎬 Real-Time Detection

Run the emotion recognition system:

python realtime_facial_analysis.py

⌨️ Controls: Press q to quit

📺 Output:

• ✅ Real-time video feed with emotion labels
• 📊 Confidence scores for each prediction
• 🎯 Face bounding boxes with tracking
• ⚡ FPS counter for performance monitoring

🧪 Training Custom Model

Open training notebook in Jupyter:

jupyter notebook model_trainign_resnet.ipynb

Contains: Complete pipeline including data loading, model training, evaluation, and export.

---

📁 Project Structure

FER2013-ResNet50-Emotion-Recognition/
│
├── 📂 models/
│   ├── FER_static_ResNet50_AffectNet.pt    # Trained model weights
│   └── training_curves.png                 # Training visualization
│
├── 🐍 realtime_facial_analysis.py          # Real-time inference script
├── 📓 model_trainign_resnet.ipynb          # Training notebook (Colab/Jupyter)
├── 📋 requirements.txt                     # Python dependencies
├── 🚫 .gitignore                           # Git ignore rules
└── 📖 README.md                            # Documentation

---

⚙️ Technical Details

🖼️ Image Preprocessing

1. Resize to 224×224 (INTER_NEAREST)
2. Convert BGR → RGB color space
3. Mean normalization: [R: 91.49, G: 103.88, B: 131.09]
4. Convert to PyTorch tensor (C×H×W)

🔧 Performance Optimizations

Optimization	Description	Benefit
FP16 Precision	Half-precision on GPU	2x faster inference
CuDNN Auto-tuning	Optimal conv algorithms	10-20% speedup
Downscaled Detection	50% resolution for face detection	50% faster detection
Threaded Capture	Non-blocking webcam stream	Eliminates I/O bottleneck
Temporal Smoothing	10-frame moving average	Stable predictions

Result: 30+ FPS on NVIDIA GPU • 10-15 FPS on CPU

---

📖 Citation

If you use this project in your research or applications, please cite:

@software{fer2013_resnet50,
  title={Real-Time Facial Emotion Recognition with ResNet50},
  author={Your Name},
  year={2024},
  url={https://github.com/yourusername/FER2013-ResNet50-Emotion-Recognition}
}

Dataset Citation

@inproceedings{goodfellow2013challenges,
  title={Challenges in representation learning},
  author={Goodfellow, Ian J and others},
  booktitle={ICONIP},
  year={2013}
}

---

📜 License

MIT License - see LICENSE file for details.

---

🙏 Acknowledgments

• FER2013 Dataset - ICML 2013 Challenges in Representation Learning
• ResNet Architecture - He et al., "Deep Residual Learning for Image Recognition"
• MediaPipe - Google's ML solutions for face detection
• PyTorch - Facebook AI Research

---

⭐ Star this repository if you find it helpful!

Built with ❤️ using PyTorch and ResNet50

Report Bug • Request Feature