EmoCNN - Emotion Recognition using Convolutional Neural Networks

EmoCNN is a straightforward project demonstrating how to perform emotion recognition on images using pre-trained Convolutional Neural Networks (CNNs). The project utilizes a dataset from Kaggle and can be easily run on Google Colab or Jupyter notebooks.

Usage

1. Download the dataset from Kaggle
2. Extract the dataset to a local directory.
3. Open notebook using Google Colab or Jupyter Notebook.
4. Update the dataset_path variable with the path to the extracted dataset folder.
5. If using Jupyter Notebook, ensure you have the Keras library installed: pip install keras.
6. Run the notebook cells to load the pre-trained model, preprocess images, and obtain emotion predictions.

Dataset

The dataset used in this project is sourced from Kaggle and contains labeled images for emotion recognition. It consists of 7 classes, and for this project, I focus on 3 classes: happy, neutral, and sad.

Dataset Link: Emotion Detection FER Dataset

Model Architecture

The model architecture utilizes a pre-trained CNN for feature extraction and emotion prediction. Initially, a 3-convolutional-layer CNN model is constructed and tested with different optimizers to find the most suitable one. In our experimentation, Adam worked best, so we built a customized model with Adam as the optimizer.

Notebooks

1. variousmodel.ipynb: Compares model behavior on different optimizers.
2. customisedmodel.ipynb: Customized model testing on test images and Grad-CAM as explainability.