Face Emotion Detection

A real-time facial emotion recognition web application trained on the FER-2013 dataset using a fine-tuned EfficientNet-B2 backbone. Detects faces in images or live webcam streams and classifies them into 7 emotions.

---

Demo

Mode	Description
Image Upload	Upload any photo — faces are detected and labeled instantly
Live Webcam	Real-time emotion detection from your browser camera

---

Emotions Detected

Angry · Disgust · Fear · Happy · Neutral · Sad · Surprise

---

How It Works

Three-stage inference pipeline:

Input Image
    │
    ▼
YOLOv8n  ──────────────────── detects person bounding boxes
    │
    ▼
OpenCV Haar Cascade ────────── detects faces within each person crop
    │
    ▼
EfficientNet-B2 ─────────────  classifies emotion (7 classes)
    │
    ▼
Labeled result with confidence score & color-coded box

---

Model Training

Trained in 5 progressive fine-tuning phases on FER-2013 (35,887 images):

Phase	Unfrozen Layers	LR	Epochs
1	Classifier head only	0.001	15
2	Last 1 backbone block	0.0001	10
3	Last 3 backbone blocks	0.00001	15
4	Last 5 backbone blocks	0.000005	15
5	Full model	0.000001	20

Techniques used:

• Class imbalance handled with inverse-frequency class weights
• Data augmentation: random flip, rotation, affine shift, color jitter, random erasing
• CosineAnnealingLR scheduler per phase
• Early stopping with best-weight restoration

---

Tech Stack

• Model: PyTorch · EfficientNet-B2 (torchvision)
• Detection: YOLOv8n (Ultralytics) · OpenCV Haar Cascade
• Web: Flask
• Training: Google Colab · FER-2013 (Kaggle)

---

Getting Started

1. Clone the repo

git clone https://github.com/Luke-Dev-Tech/Emotion_Detection_Computer_Vision.git
cd Emotion_Detection_Computer_Vision

2. Install dependencies

pip install -r requirements.txt

3. Add the model weights

Place FaceEmotionDection_argumented_final.pth inside the EmotionDection/ folder.

4. Run the app

cd EmotionDection
python app.py

Open http://localhost:5300 in your browser.

---

Project Structure

├── FACE_EMOTION.ipynb          # Full training notebook (Colab)
├── requirements.txt
├── EmotionDection/
│   ├── app.py                  # Flask app
│   └── templates/
│       ├── index.html          # Image upload page
│       └── livecam.html        # Real-time webcam page

---

Dataset

FER-2013 — 35,887 grayscale 48×48 face images across 7 emotion classes.

---

License

MIT