A personal project leveraging both Statistical Learning and Deep Learning techniques to accurately recognize and classify faces.

PyTorch and polars are used for efficiently handling the data in a distrubuted fashion.

The faces94 dataset is utilized to train and evaluate models.

faces94 consists of thousands of images, concretely around 20 images from more than 300 distinct individuals.

A facial recognition system is built with two well-known Statistical Learning methods implemented from scratch: Eigenfaces and Fisherfaces. This system is able to discern whether a person belongs or not to the given database as well as its identification in case of the former.

Furthermore, an AI model (Convolutional Neural Network) is trained replicating the LeNet-5 architecture using PyTorch as the preferred Deep Learning framework.

Check the Jupyter Notebook showcasing the models with additional information about the algorithms.

The following software is needed to successfully run the project:

• Python >= 3.11
• pip >= 23.3.2

Optionally, if you own a GPU it is recommended to have CUDA installed.

For downloading and automatically managing the images it is necessary to use a Unix-based OS like Linux or macOS. A script that generalizes for Windows is being worked on.

Python libraries can be directly downloaded after creating a virtual environment with:

python3 -m pip install -r requirements.txt

⚠️ All Python scripts must be run from the speficic directory they are located ⚠️

Firstly, it is needed to download and sort the images into three partitions:

python3 set_up_data.py

This will create a data folder with all the relevant images split into train, validation and test partitions.

After this you can train the models, for example, if we want to train the CNN:

python3 train_lenet.py

FacialRecognizer/
├── config/
│   ├── .env
│   └── url.py
│
├── controllers/
│   ├── data_extractor.py
│   └── data_images.py
│     
├── data/
│   ├── train/
│   ├── validation/
│   └── test/
│
└── set_up_data.py

The config folder simply stores the URL for the dataset within a variable in url.py. Since the URL was long, it was kept in an environmental variable.

Within controllers, the main data extraction and partition process is implemented: set_up_data.py serves as a main file for running the different methods of the DataExtractor class in data_extractor.py. data_images.py turns each image into an array of pixels stored in a polars dataframe for later processing.

Concretely, faces94 is downloaded and a sample of females is guaranteed to be present in all datasets as to make sure the models generalize well. As to simulate the presence of individuals who do not belong to the original database, certain images are hidden from training in the folders validation and test.

FacialRecognizer/
├── stats/
│   ├── pca.py
│   ├── lda.py
│   ├── knn.py
│   ├── eigenfaces.py
│   └── fisherfaces.py
│
├── train_eigenfaces.py
└── train_fisherfaces.py

pca.py includes the code for applying Principal Component Analysis (PCA) for dimensionality reduction.

lda.py implements Fisher's Linear Discriminant Analysis (LDA) method for maximizing the spread between classes and minimizing the within-class variance.

A custom k-Nearest Neighbors (k-NN) algorithm can be found in knn.py. It chooses the most voted class among the selected candidates. The user may opt for one of the following statistical distances: 'cosine', 'seuclidean', 'euclidean', 'canberra'.

eigenfaces.py and fisherfaces.py define linear discriminants that feed the classifier defined in knn.py.

Both scripts train_eigenfaces.py and train_fisherfaces.py build a facial recognizer system based on the Eigenfaces and Fisherfaces approaches, tuning the hyperparameters with Bayesian Optimization.

FacialRecognizer/
├── deep_learning/
│   ├── dataset.py
│   └── lenet.py
│
└── train_lenet.py

dataset.py sets up a custom PyTorch Dataset structure to store the data and lenet.py lays out the code for defining LeNet-5 core architecture and its learning process. Finally, train_lenet.py fits a CNN to the data and performs classification, procuring the accuracy on test images.