Face Recognition—Eigenfaces

Abstract

The aim of the project is to recognize a person’s face by making a comparison between its characteristics and those of other people’s faces which are already known. To perform this task, Eigenfaces method is implemented together with a previous preparation of the images. Initially, outlier images are identified within a face and landscape dataset, using metrics such as Manhattan, Euclidean, Chebyshev and Minkowsky distances $\left( {p = \frac{5}{2}\,{\text{and}}\,p = \frac{\sqrt 2 }{2}} \right)$. Then the results are compared to determine which one of the metrics has better performance in the identification of rare images. In total 2.470 images (2.260 faces and 210 natural landscapes) were used. Secondly, the methodology of Eigenfaces was implemented with a total of 3059 images of faces and then, by distance measurements, distances of new images projected on the subspace were compared with those used to form the Eigenfaces. Thirdly, a lineal discriminant analysis (LDA) was used on the subspace generated through PCA. Finally, for the prediction of faces of women, men and landscapes, supervised and unsupervised classification methods were used and compared. For the supervised classification, the Multinomial Logistic Regression and Linear Discriminant Analysis (LDA) were used, and the unsupervised analysis was done using K-means, T-Distributed Stochastic Neighbor Embedding (TDSNE) and Agglomerative Clustering.