Author: José David Bermúdez Castro
Original title: Estimativa da Idade a partir de Imagens Faciais
Language: Brazilian Portuguese
This thesis aims to investigate methods for age estimation from facial images. The impact of distinct factors over the estimate’s accuracy is assessed, specifically the accuracy in the location of face fiducial points, feature extraction and dimensionality reduction methods, and regression techniques. Additionally, the dependence on race and gender in the accuracy of age estimation is assessed. Five performance metrics have been considered: the mean absolute error (MAE), the mean absolute error per decade (MAE / D), the mean absolute error for age (MAE / A), the cumulative score (CS) and confidence intervals (CI). The experiments were performed using two public databases, whose images are labeled with the age of the face. The results showed the impact of the automatic method for detection of fiducial points of the face has a moderate impact on the accuracy of the estimates. Among the analyzed variants, the one with the best accuracy was the system that employs the Active Appearance Models (AAMs) as feature extraction method, the Principal Components Analysis (PCA) as dimensionality reduction method, and Support Vector Regression (SVRs) as a technique to do regression.