Fine Tuning Age-estimation with Global and Local Facial Features

This paper proposes an advanced age-estimation approach that combines global and local features derived from a facial image. Active Appearance Models (AAMs) technique is used to construct the global facial features, while local facial features are generated from Local Binary Pattern (LBP) encoding. Ageestimation is performed in a two-step method: coarse (initial) prediction followed by a refining step (fine prediction). Support Vector Regression (SVR) is used to learn the initial and refined aging function. In this work, we determine the efficacy of combining these approaches against a standard face database for age-estimation, FG-NET, and a newly developed face database on Southeast Asians, Vietnamese Longitudinal Face (VLF) database. Further, this work will provide some insight into which type of features, global or local, are better and how to combine them to improve age-estimation. Standard performance measures of mean-absolute error (MAE) and cumulative score (CS) are used to evaluate the two approaches. Although the results against FGNET are not industry best of 4.04 yrs [1] and 3.96 yrs [2], overall MAE and MAE per decade indicate that this approach to fine tuning age-estimation demonstrates promise.