Hierarchical age estimation with dissimilarity-based classification

This paper proposes a novel approach that models the process of aging using Active Appearance Models (AAMs) and Ensemble of Classifiers for Age Estimation. The approach treats the problem of age estimation as a combination of classification and regression problems. In this approach, face image is encoded using the statistically driven AAMs which uses both shape and appearance models to form a combined model to represent the face image as a feature vector. A global classifier is then used to obtain a rough idea about the age by distinguishing between child/teen-hood and adulthood, while final age estimation is made using regression functions. To reduce misclassification error, an ensemble containing various classifiers trained on multiple dissimilarities has been used. The images thus classified are passed on to different aging functions for further accurate age estimation.Experiments have been performed on the publicly available FG-NET database and the Center for Vital Longevity Face Database to test the approach. It has been observed that the proposed approach has the lowest Mean Absolute Error (MAE) and the highest Cumulative Score when compared with other published results. It is further tested on IIT Kanpur database consisting of images of age group 18–34 acquired under semi-controlled environment.