Biometric features estimation from facial images
Abstract: In this work we propose a two novel approaches for estimating biometric features based on face images. We estimate age, race and gender, but focus mainly on age estimation. Our first approach improves the age estimation by formulating the Diffusion Framework as a boosted supervised learning algorithm, where at each step we estimate a subset of the biometric attributes that are used to augment the feature space in the succeeding estimation step. The feature spaces are embedded in the Diffusion embedding space to yield adaptive bases that are used for age regression. In our second contribution we propose to represent the set of face image by a deep tree, partitioned according to biometric attributes and coarse age estimates. Thus, the image data is in each graph leaf relates to simpler manifold geometries and can be estimated separately. The tree construction and regression are based on Kernel SVM regressors and we show that their accuracy can be significantly improved by applying low rank distance learning. We tested the resulting schemes on the state-of-the-art MORPH-II dataset, and showed a major improvement compared to contemporary works.
An MSc, thesis supervised by Prof. Yosi Keller