Facial Age Estimation By Ordinal Regression

Facial age estimation is a task to automatically estimate the age of human face based on a facial image.Due to its broad application prospect,facial age estimation has attracted increasing attention in the field of Pattern Recognition,Multimedia Com-puting,Human Machine Interaction,etc..Previous works usually designed facial age estimation methods on the basis of traditional regression or classification formulations.Considering the fact that age labels imply a natural total order relationship and misclas-sifying a face to an age which is far away from its true age is unbearable,this paper proposes to formulate this task as an ordinal regression problem and achieves the fol-lowing main contributions:(1)We propose an approach named Harmonious Ordinal Regressions(HOR) which establishes a connection between two traditional ordinal regression approaches,Threshold Model and Binary Decompositions.By choosing a proper trade-off,HOR inherits the fast estimating speed of Threshold Model and the high accuracy of Binary Decompositions.Experiments on three datasets,FG-NET,MORPH and ChaLearn,show that HOR achieves accuracy comparable to that of Binary Decom-positions and estimating speed similar with that of Threshold Model.Experiments also show that HOR outperforms state-of-the-art age estimation methods.(2)We propose Total Ordering Preserving Projection(TOPP)to identify an subspace which can preserve the total order structure to the best.By requiring the projection to distinguish high-order examples and low-order ones in all partitions induced by ordinal relationship,TOPP utilizes the total order information naturally and identi-fys projections which can present a clear aging trend in the task of facial age esti-mation.Experiments on three benchmark datasets show that replacing the original input space with the feature space learned by TOPP can improve the accuracy of SVM-HOR significantly.