Facial Sentiment Analysis using Convolutional Neural Network (CNN)

📌 Table of Contents

• Project Background
• Key Features
• Technologies Used
• Installation
• Usage
• Detailed Results
• Future Work
• Acknowledgments
• License

---

🌟 Project Background

Year: 2022-2023 (Engineering Internship)
Supervisor: Mr. Bassem Hmissa (LEONI WIRING SYSTEMS)
Institution: National School of Electronics and Telecommunications of Sfax, Tunisia

🎯 Motivation

Commercial APIs (e.g., Microsoft Face API) treat facial analysis as a "black box," limiting customization and interpretability. This project was developed to:

• Demystify AI models by building face detection, emotion recognition, and facial identification systems from scratch.
• Analyze hyperparameter impact (epochs, batch size, optimizers) on model accuracy.
• Provide a flexible SaaS solution adaptable to client-specific datasets.

---

🚀 Key Features

1. Face Detection

• Haar Cascade: Real-time detection using OpenCV.
• YOLOv4: Higher accuracy (94.82%) but sensitive to face angles/lighting.
  • Limitations: Struggles with rotated faces () and color resemblance (.

2. Emotion Detection

• 7-class CNN (angry, disgust, fear, happy, neutral, sad, surprise).
• Dataset: FER-2013 (35,887 images, 48x48px).
• Best Model:

  Model.add(ZeroPadding2D())  
  Model.add(Conv2D(32, (3,3), activation='relu'))  
  Model.add(MaxPooling2D())  
  Model.add(Dense(7, activation='softmax'))  

3. Facial Recognition

• Dataset: LFW (13,233 images) + custom-injected faces.
• Custom CNN: Achieved 76.34% training accuracy.

---

🛠️ Technologies Used

Category	Tools/Libraries
Frameworks	TensorFlow, Keras, Darknet (YOLOv4)
Computer Vision	OpenCV, PIL
Data Processing	NumPy, Pandas, Matplotlib
Environment	Google Colab, Kaggle API

---

📦 Installation

1. Clone the repository:

   git clone https://github.com/mariam-khediri/Facial-Sentiment-Analysis-using-CNN.git  
   cd facial-sentiment-analysis  

2. Install dependencies:

   pip install -r requirements.txt  # Sample requirements.txt included below  

   requirements.txt:

   tensorflow==2.8.2  
   opencv-python==4.6.0  
   numpy>=1.21.0  
   matplotlib>=3.5.0  
   

3. Download pretrained models:
   • YOLOv4 weights
   • Haar Cascade XML

---

🖥️ Usage

Real-Time Emotion Detection

python emotion_detection.py --mode=camera  

Output:

• Processes webcam feed → detects faces → predicts emotions ( ).
• Supports image input: --image=test.jpg.

Facial Recognition

python facial_recognition.py --dataset=lfw --epochs=120  

Output:

• Recognizes faces from LFW dataset or custom-injected images

---

📊 Detailed Results

Emotion Detection (CNN)

Hyperparameter	Best Value	Accuracy (Train/Test)
Optimizer	Adam (lr=0.001)	94% / 51%
Epochs	30	94.75% (Train)
Batch Size	64	47.38% (Test)

Key Observations:

• Higher epochs → overfitting (test loss ↑ to 6.21).
• SGD optimizer underperformed (max 25.83% test accuracy).

Facial Recognition (Custom CNN)

Batch Size	Epochs	Train Accuracy	Test Accuracy
64	120	76.34%	14.32%
32	25	32.37%	10.34%

Challenges:

• Low test accuracy due to dataset complexity (5,749 identities).
• Dropout layers reduced overfitting ( vs ).

---

🔮 Future Work

• Data Augmentation: Improve test accuracy with transformations (rotation, scaling).
• Transformer Models: Experiment with Vision Transformers (ViT).
• API Deployment: Flask/Django backend for SaaS integration.

---

🙏 Acknowledgments

• Professional Supervisor: Mr. Bassem Hmissa (LEONI).
• Dataset Providers: Kaggle (FER-2013, LFW).
• Institutional Support: Sfax University, Tunisia.

---

📜 License

MIT © Mariem Khedhiri