This repository contains the various AI and interactive graphics project files.
First Exploratory Data Analysis will be done on the dataset, then various convolutional neural networks capable of performing facial expression recognition will be implemented and their performances will be compared.
To run the files from this repository you need to add this folder to your Google Drive.
The folder contains the dataset (which can be downloaded from here) and the files relating to the various previously trained CNNs.
Exploratory data analysis is an essential tool for analyzing the dataset and understanding its peculiarities.
Thanks to it, it's possible to grasp important aspects of the dataset such as imbalances between classes, missing values, etc.
The Exploratory_Data_Analysis.ipynb file implements the EDA on the dataset, displaying among other things a sample of images and the histogram of the classes, which allows us to understand how these are distributed.
In particular, looking at the generated histogram of the classes, it can be seen that the "disgust" class has far fewer images than the other classes, so it will be necessary to do data augmentation to balance the dataset and efficiently train the implemented models.
The Five-Layers-CNN.ipynb file implements a 5-layers convolutional neural network (whose structure is described in this pdf).
The learning curves related to accuracy and loss are shown below:
After a tuning phase on the dropout rate, leaning rate, L2 regularization rate and batch size, the following results were achieved:
training accuracy: 0.6821 | validation accuracy: 0.5602 | test accuracy: 0.5548
The confusion matrix is shown below:
Detailed analyzes of the confusion matrix, learning curves and classification report are available in the .ipynb file.
The file VGG16.ipynb implements a VGG16 (structure of which is described here).
After a network tuning phase and after making some changes in the net structure, the following results were achieved:
training accuracy: 0.7718 | validation accuracy: 0.6217 | test accuracy: 0.60
The graphs of the learning curves and the confusion matrix are shown below.
A detailed analysis of these and of the classification report is contained in the .ipynb file.
