CMPE255

Facial Emotion Recognition

This is a Kaggle challenge. The dataset consists of 48x48 grayscale images of faces. All the faces are centrally aligned and occupy around the same amount of space. The objective of this project is to recognise the emotion depicted by the human in the image.

We have built multiple models using Convolutional Neural Networks, Support Vector Machines, Recurrent Neural Networks, Decision trees and K-Nearest Neighbours. The notebooks are present in the src folder. Dataset can be found in here.

Steps to run

Follow the folder Structure

├── src
|    └── notebook.ipynb
└── Data
     └──  fer2013.csv

• Download the dataset from here
• Download the notebook of interest from here
• The commands to install the dependencies are provided in each notebook. Run the commands to install the necessary dependecies
• Assign the path to the src folder to the variable 'directory' present in the notebooks
• Run the notebook

Preprocessing techniques:

• Min-max Normalization
• Principal Component Analysis
• L2 Normalization

Feature Engineering techniques:

• Bag of Visual words
• Vector of Locally Aggregated Descriptors

Algorithms Used:

• Convolutional Neural Networks
• Support Vector Machines
• Recurrent Neural Networks
• Decision tree
• K-Nearest Neighbours
• Logistic Regression

Top five accuracies in the Kaggle leaderboard

• 71.16%
• 69.26%
• 68.82%
• 67.48%
• 65.25%

Accuracies achieved by us:

Algorithm	Accuracy
Convolution Neural Networks(Data augmentation)	64.35%
Convolution Neural Networks	55.84%
Logistic Regression.	38.49%
Recurrent Neural Networks	38.35%
KNN	35.52%
Support Vector Machine	33.18%
Decision Tree	25.09%