Facial Age Estimation Based On Adaptive Multivariate Multiple Regression

Face images can intuitively reflect many personalized characteristics related to human beings,not only including biological characteristics,such as gender,age,race and skin color,but also convey rich and diverse attitudes and emotions.Among these personalized features,age is a very useful biological information.Given a real face image,an age estimation model is established on the basis of extracting age-related facial features,which can automatically output the predicted age of face image.This learning process based on computer vision is called facial age estimation.To solve the problem of insufficient and incomplete training data of face images,facial age estimation based on label distribution learning generates an age label distribution which covers a certain number of class labels for each facial image,then utilizes label distribution model to estimate facial age.Label distribution represents the description degree of each label to the corresponding instance,and the description degree is higher when the label is closer to the chronological age.Traditional label distribution learning,which supposes facial aging has the same tendency at different ages,generates Gauss distribution centered at the chronological age with a unified standard deviation for each face image.However,the trend of facial aging is significantly different at different ages,e.g.,the facial appearance during childhood and senior age generally changes faster than that during middle age.As a result,it is unreasonable to set a unified standard deviation for all ages.In order to address the problem that multivariate multiple regression based traditional label distribution learning methods cannot generate the label distribution according to the tendency of facial aging,we proposed a facial age estimation method based on adaptive multivariate multiple regression.The proposed method generates the discrete Gaussian distributions with different standard deviations adapted to different ages,and it uses partial least square model to effectively utilize adjacent facial ageing information to predict facial age.The adaptive multiple regression method can adaptively establish the label distribution with smaller standard deviation for the age with larger changes in facial features,and it also establishes the label distribution with larger standard deviation for the age with more stable facial features.So the adaptive multiple regression method can make more effective use of the aging information of the neighbor's age in the age estimation model.The experimental results on MORPH database show that the proposed method can improve the accuracy of face age estimation algorithm compared with the existing traditional facial age estimation methods based on label distribution learning.