Reconhecimento de faces via PCA: análise de desempenho

Abstract

With the augment of criminal activities as bank robbery, kidnapping, etc., the needs of identifying a person in a safe manner became indispensable. Several algorithms have been proposed with the aim of resolving some of the challenges found in the retrieval systems as pose, illumination, age and facial expression effects. A very well diffused method in face recognition is the PCA - Principal Component Analysis, proposed by Turk and Pentland in 1991. The PCA method has as the main goal to seek the principal components of faces distribution, which means, the eigenvectors of the covariance matrix of the face set, where each eigenvector represents the variation among the faces (auto faces) that can be seen as a feature set. This work dedicates to the performance of the PCA method in face retrieval, taking into consideration the use of different poses, facial expression and illumination effects. The PCA is also evaluated when the query face has some of its features modified (i.e. noise, occlusion, cutting, etc). An analysis is done taking into consideration how these modified features could change the retrieval results, besides the illumination and poses influence. The performance of PCA method is also evaluated considering which ones and the quantity of eigenvalues to be selected by the algorithm, with the aim of reaching good retrieval results. With the aim of verifying the illumination problems, a smooth filter is used in a set of faces and then the PCA method is applied in this set of images considering only those images pieces that have texture and noise presence, derived of the difference between the original and the smoothed image. The main goal is to improve the face retrieval under several illumination conditions.