Facial Age Estimation For Unconstraint Images

Age estimation aims to estimate an accurate age of a given face image,which is an important part of facial attribute estimation.Nowadays,the precision of age estimation has been greatly improved for the rapid development of convolutional neural network.Many researchers try to find a more suitable age encoding or loss function to help the network learn better about age estimation task.Similar to them,we want to solve some existing issues of current works to make our contribution to this topic.The main contributions of this dissertation include the follow aspect:Firstly,we propose a new age encoding named soft-ranking.Current age encoding methods are unable to accurately describe the property of aging process,which prevent the network from learning age estimation task better.To eliminate this problem,softranking encoding is designed.By enabling the encoding method to simultaneously depict the correlation of adjacent age and ordinal information of aging,the convolutional network can grasp the prior knowledge of human being and thus make accurate estimation.The experiments results show that soft-ranking outperforms both label distribution learning and ranking,the other two popular encoding methods in age estimation.Secondly,an automatic age encoding method based on label distribution learning is put forward.We investigate the effect of the parameter in label distribution learning and find out that an appropriate value not only help the network to distinguish indistinct difference in adjacent ages but also to catch the correlation information of aging.So,we innovatively combine the label distribution learning with variance loss function together for automatedly finding a proper label distribution value.Sufficient experimental verification is carried out to show the effectiveness of our proposed method.Finally,we add some auxiliary regularization tasks to the help the network in eliminating the overfitting issue.By forcing the network to conduct age estimation task on masked feature maps,the network is encouraged to pay attention to the whole feature maps and thus able to make reliable estimation.The experiment results show that our method bring extra improvements when compared with baseline.