Face Recognition using PCA, LDA, and Autoencoder - 2022

Purpose

Testify the effectiveness of current popular machine learning methods (PCA, LDA, and Autoencoder) in face recognition task. 

Dataset

FacePix Database from ASU.

Procedure

1. Vectorize the images in the training set.

2. Center the mean and normalize each image vector.

3. Compute the covariance and retain p eigenvectors.

4. Project the training set into the p–dimensional subspace.

5. Use the Nearest Neighbor (NN) for classifications.

Results

• LDA performed the best, PCA performed the second best, and Autoencoder was third.

• All methods increased accuracy as the specified number of dimensions increased and eventually achieved 100% accuracy.

Some notes on AutoEncoder

• Autoencoder is a deep-learning-based approach for dimensionality reduction. 

• The idea of an autoencoder is to create a compressed-image representation and convert it back to its original form. 

• The neural learning is set to find the best reduced-dimensionality representation so that the error between the original and the reconstruction is the smallest.

Some notes on Eigenvectors

• PCA: p eigenvectors are directly the eigenvectors of the covariance matrix.

• LDA: the eigenvectors are the function of the within-class scatter matrix and the between-class scatter matrix . 

Compute PCA

Compute LDA