Emotion Detection Using Convolutional Neural Networks (CNN)

Overview

This project aims to classify human facial emotions into one of seven categories using deep Convolutional Neural Networks (CNN). The model is trained on the FER-2013 dataset, which was unveiled at the International Conference on Machine Learning (ICML). The dataset comprises 35887 grayscale facial images of size 48x48, representing seven emotions: anger, disgust, fear, happiness, neutral, sadness and surprise.

Dependencies

• Python 3, OpenCV, Tensorflow
• To install the required packages, execute: pip install -r Requirements.txt.

Basic Usage Instructions

The repository is compatible with tensorflow and utilizes the Keras API via tensorflow.keras.

• Clone the repository and navigate to the directory:

git clone https://github.com/hoangtung719/Emotic-Detection.git
cd Emotic-Detection

• If you want to retrain the model, download the FER-2013 dataset from here and then save it to the Source_code directory.

• To train the model:

cd Source_code
jupyter notebook Model_training.ipynb

Execute all cells within the Model_training.ipynb file.

• To view predictions without retraining, download the pre-trained model from here and:

cd Source_code
jupyter notebook Model_testing.ipynb

Execute all cells within the Model_testing.ipynb file.

• Directory structure:

Source_code:
 - Model_training.ipynb (file)
 - Model_testing.ipynb (file)
 - haarcascade_frontalface_default.xml (file)
 - BeVietnamPro-Regular.ttf (file)
 - Emotion_recognition_model.h5 (file)

• By default, this implementation detects emotions on all faces from webcam feed. With a simple 4-layer CNN, an accuracy of 65.97% was achieved after 60 epochs.

Data Preparation (Optional)

• The original FER-2013 dataset on Kaggle is available as a csv file. I have used a pre-converted image version for training/testing.

• For experimenting with new datasets, simply change the paths in train_dir and val_dir.

Algorithm

• Faces are detected using the haar Cascade method from webcam feed.

• Detected faces are resized to 48x48 and converted to input for CNN.

• The network outputs a list of softmax scores for seven emotion categories.

• The emotion with the highest score is displayed on screen.

• The BeVietnamPro-Regular.ttf font is used to display Vietnamese emotions on screen.