Research On The Age Estimation Method Of Face With Orderly Labels

As an important biological feature reflected by human faces,age is widely applied in many fields such as electronic commerce information management,security monitoring and so on,which can be simply deduced from the appearance.Age estimation based on facial images has recently become an active topic in the field of computer vision and human-computer interaction.Facial age estimation is a process of using the machine learning algorithms to automatically estimate the approximate age or age range of the object.Although the facial age estimation has made a lot of progress in recent years,how to find the appropriate feature representation and estimation model are still a challenge to this problem.This is mainly due to the uncontrollable and personalized characteristics of the aging process,and the lack of sample images in the facial age database.Traditional dimensionality reduction and age estimation algorithms usually only consider the differences between age groups,while ignoring the ordered structure between age labels.However,this ordered structure tends to better depict the relationship between age labels.Therefore,this paper focuses on the influence of the ordered structure between age labels on the performance of facial age estimation.The results show that the facial age estimation which considering the ordered structure can obtain better performance.The main contents of this paper are as follows:(1)Two label-sensitive dimensionality reduction algorithms are proposed,which are based on Supervised Neighborhood-based Fisher Discriminant Analysis and Maximum Marginal Criterion respectively.In order to ensure that the ordered structure between age labels can be fully excavated in the reduced dimensionality space,the weight of sample pairs is assigned according to the feature similarity and difference degree between age labels.(2)Considering the complicated facial aging process and the local statistical properties of the reduced feature,a two-step local regression algorithm is proposed for facial age estimation.For a sample image,its nearest Euclidean neighbors are searched firstly,and then a Label Distribution Support Vector Machine regressor is trained based on these neighbors.Instead of applying all samples,local regression based on neighboring samples,not only simple,saving time and simplifying the computational load,and can also generate comparably accurate ageing decision functions.In this paper,two label-sensitive dimensionality reduction algorithms and a two-step local regression algorithm are proposed,which effectively utilize the ordered information between age labels.To some extent,this method can make up for the shortage of the number of facial images in the database.The validity of proposed algorithm in this paper is verified in the open FGNET database.