Machine learning: Facial Expression Recognition Project

By Beni Tibi, and Chen Ben Ami
Ariel University

---

Overview

This repository showcases a facial expression recognition system using deep learning. The system compares four main approaches:

1. Decision tree model
   • A simple baseline that always predicts the most frequent class.
2. Random Forest
   • An ensemble of decision trees for improved performance.
3. Basic Fully Connected Network (MLP)
   • A simple neural network flattening the image and using fully connected layers.
4. Advanced Convolutional Neural Network (CNN)
   • Multiple convolutional blocks (Conv2D, BatchNorm, ReLU, MaxPool, Dropout) for superior performance.

Each model is trained to recognize 7 emotion classes (angry, disgust, fear, happy, neutral, sad, surprise) from grayscale images sized 48×48. After training, we demonstrate a live inference script where the advanced CNN classifies expressions in real time via a webcam feed.

---

Purpose

1 Comparison

• Evaluate accuracy, precision, recall for each approach.
• Show that the CNN typically outperforms simpler models on an image-based problem.

2 Practical Flow

• Data Download (from Kaggle)
• Data Preparation (resize, grayscale, normalization)
• Training multiple models
• Evaluation (metrics)
• Live Inference using the advanced model.

---

Dataset

• Source: Kaggle - Face Expression Recognition Dataset
• Contains 7 categories in subfolders: e.g., train/angry/, train/happy/, etc.
• Images are manually split into train/, validation/, and test/ subfolders.

---

Installation Instructions

1. Clone the repository (or download):

   git clone https://github.com/YourUserName/facial_expression_recognition_project.git
   cd facial_expression_recognition_project

2. (Optional) Create a virtual environment:

   python -m venv venv
   source venv/bin/activate   # or venv\Scripts\activate on Windows

3. Install required libraries:

   pip install -r requirements.txt

   • This includes PyTorch, NumPy, Matplotlib, OpenCV, scikit-learn, kaggle, etc.

4. Kaggle Setup (if you need to download the dataset):

   • Place your kaggle.json credentials in ~/.kaggle/ (Linux/Mac) or %HOMEPATH%\.kaggle\ (Windows).
   • Ensure correct permissions (e.g., chmod 600 ~/.kaggle/kaggle.json).

5. Download dataset (if required):

   python download_dataset.py

   • This will fetch the Face Expression Recognition dataset and unzip it under data/face-expression-recognition-dataset/.

6. Prepare the .pt data (preprocessing):

   python dataset_preparation.py

   • Reads raw images from train/, validation/, test/, converts them to (N,1,48,48) shape, normalizes, and saves .pt files to processed_data/.

---

How to Run the Models

1. Decision Tree

   python decision_tree.py

   • Trains the baseline model, logs classification report, saves results in decision_tree_results.txt.

2. Random Forest

   python random_forest.py

   • Trains the ensemble model, logs classification report, saves results in random_forest_results.txt.

3. Basic NN (MLP)

   python basic_nn.py

   • Multi-layer perceptron, logs classification report, saves results in basic_nn_results.txt.

4. Advanced CNN

   python advanced_network.py

   • Trains the deeper convolutional network, saves best weights to advanced_model.pth.

5. Live Inference

   python live_inference.py

   • Must have advanced_model.pth from the advanced CNN.
   • Opens webcam, detects face, classifies expression in real time. Press q to quit.

---

Contact / Credits

• Authors:
  • Chen Ben Ami
  • Beni Tibi
• Institution: Ariel University

Kaggle Reference:
Face Expression Recognition Dataset

---

Future Ideas

1. Data Augmentation (random flips/rotations) to improve generalization.
2. Class Weights or oversampling for underrepresented expressions (e.g., disgust).
3. Transfer Learning from pretrained networks (like ResNet) for potentially higher accuracy.
4. Hyperparameter Tuning (learning rate, dropout rates, layer dimensions).

---