🎭 Facial Emotion Detection System

A real-time facial emotion detection system using Convolutional Neural Networks (CNN) trained on the FER-2013 dataset.

Classifies 6 emotions: 😠 Angry · 😨 Fear · 😄 Happy · 😢 Sad · 😲 Surprise · 😐 Neutral

---

📋 Table of Contents

• Overview
• Emotion Classes
• Project Structure
• Preprocessing Pipeline
• Model Architecture
• Setup & Installation
• How to Run
• Team

---

🔍 Overview

This project uses deep learning to detect and classify facial emotions in real-time through a webcam feed. It was developed for the ICT3212 - Introduction to Intelligent Systems module at Rajarata University of Sri Lanka.

Feature	Details
Dataset	FER-2013 (Kaggle)
Model	Custom CNN (4 Conv blocks)
Input	48×48 grayscale images
Classes	6 emotions
Interface	Streamlit web app
Detection	OpenCV Haar Cascade

---

🎯 Emotion Classes

Label	Emotion	Samples	Percentage
0	😠 Angry	~4,953	14.0%
1	😨 Fear	~5,121	14.5%
2	😄 Happy	~8,989	25.4%
3	😢 Sad	~6,077	17.2%
4	😲 Surprise	~4,002	11.3%
5	😐 Neutral	~6,198	17.5%
	Total	~35,340	100%

Note: The original Disgust class was removed due to insufficient samples (~547).

---

📁 Project Structure

Facial-Emotion-Detection-System/
│
├── 📓 Notebooks
│   ├── Facial_Emotion_Detection.ipynb    # Full pipeline (train + inference)
│   ├── Dataset_Exploration.ipynb         # Dataset organization & analysis
│   └── Image_Preprocessing.ipynb         # Preprocessing steps
│
├── 📂 src/
│   ├── data_preprocessing.py             # Data loading & processing
│   ├── model.py                          # CNN architecture definition
│   ├── train.py                          # Training script
│   └── app.py                            # Streamlit real-time app
│
├── 📂 data/
│   ├── fer2013.csv                       # FER-2013 dataset (download separately)
│   └── organized/                        # Class-wise image folders (auto-generated)
│       ├── Angry/
│       ├── Fear/
│       ├── Happy/
│       ├── Sad/
│       ├── Surprise/
│       └── Neutral/
│
├── 📂 Models/                            # Saved trained models
├── 📄 requirements.txt                   # Dependencies
├── 📄 Model_Card.md                      # Hugging Face model card
├── 📄 Dataset_Exploration_Report.txt     # Dataset analysis report
├── 📄 Image_Preprocessing_Report.txt     # Preprocessing report
├── 📄 Project_Report.txt                 # Full project report
└── 📄 README.md

---

⚙️ Preprocessing Pipeline

 ┌─────────────┐     ┌─────────────┐     ┌─────────────┐     ┌─────────────┐
 │  1. RESIZE  │────▶│ 2. NORMALIZE│────▶│ 3. TENSOR   │────▶│  4. SPLIT   │
 │  48 × 48    │     │  [0, 1]     │     │ (N,48,48,1) │     │ 80/20       │
 └─────────────┘     └─────────────┘     └─────────────┘     └─────────────┘

Step	Operation	Before	After
1	Resize	Raw CSV pixels	48 × 48 images
2	Normalize	[0, 255] uint8	[0.0, 1.0] float32
3	Convert to Tensor	NumPy (N, 48, 48)	TF Tensor (N, 48, 48, 1)
4	Train/Test Split	35,340 total	28,272 Train / 7,068 Test

---

🧠 Model Architecture

Input (48 × 48 × 1)
       │
       ▼
┌──────────────────────┐
│ Conv2D(64) + BN + MP │──── Block 1
│ + Dropout(0.25)      │
└──────────┬───────────┘
           │
┌──────────▼───────────┐
│ Conv2D(128) + BN + MP│──── Block 2
│ + Dropout(0.25)      │
└──────────┬───────────┘
           │
┌──────────▼───────────┐
│ Conv2D(512) + BN + MP│──── Block 3
│ + Dropout(0.25)      │
└──────────┬───────────┘
           │
┌──────────▼───────────┐
│ Conv2D(512) + BN + MP│──── Block 4
│ + Dropout(0.25)      │
└──────────┬───────────┘
           │
    ┌──────▼──────┐
    │   Flatten   │
    └──────┬──────┘
           │
  ┌────────▼────────┐
  │ Dense(256) + BN │
  │ + Dropout(0.5)  │
  └────────┬────────┘
           │
  ┌────────▼────────┐
  │ Dense(512) + BN │
  │ + Dropout(0.5)  │
  └────────┬────────┘
           │
    ┌──────▼──────┐
    │ Dense(6)    │
    │ Softmax     │
    └─────────────┘

Training Configuration:

• Optimizer: Adam (lr = 0.0001)
• Loss: Categorical Crossentropy
• Callbacks: ModelCheckpoint, ReduceLROnPlateau, EarlyStopping

---

🚀 Setup & Installation

Prerequisites

• Python 3.8+
• Webcam (for real-time detection)
• FER-2013 Dataset

Installation

# 1. Clone the repository
git clone https://github.com/Tharinda-Pamindu/Facial-Emotion-Detection-System.git
cd Facial-Emotion-Detection-System

# 2. Create virtual environment
python -m venv .venv

# 3. Activate virtual environment
# Windows:
.venv\Scripts\activate
# macOS/Linux:
source .venv/bin/activate

# 4. Install dependencies
pip install -r requirements.txt

# 5. Download FER-2013 dataset and place it in:
#    data/fer2013.csv

---

▶️ How to Run

Option 1: Jupyter Notebooks (Recommended)

jupyter notebook

Run the notebooks in order:

1. Dataset_Exploration.ipynb — Explore & organize the dataset
2. Image_Preprocessing.ipynb — Preprocess images
3. Facial_Emotion_Detection.ipynb — Train model & run detection

Option 2: Python Scripts

# Train the model (20-60 min depending on hardware)
python src/train.py

# Run the Streamlit app
streamlit run src/app.py

Option 3: Batch Scripts (Windows)

# Train
train_model.bat

# Run app
run_app.bat

---

👥 Team — Inflators

Member	Role	Github
DTPD Wickramasinghe	👑 Group Leader	Tharinda
DVTR Vitharana	Member	Thinuka
RSR Ranathunga	Member	Sanka
DDSS Kumasaru	Member	Dilakshi
SHD Mihidumpita	Member	Hansa

---

ICT3212 - Introduction to Intelligent Systems Rajarata University of Sri Lanka