Age Estimation Method Of Face Images Based On Multi-task Learning

Face image age estimation is a basic research,which has a wide range of applications in social media,security monitoring and other fields.Age estimation tasks are different from common image classification tasks because of the correlation and order between ages,the features required for age estimation are similar to the facial features of their adjacent ages,and there is ambiguity between labels;age estimation has attribute correlation.The recognition of other face attributes has a certain influence on age estimation;age estimation requires not only global features,but also age-sensitive features.Aiming at the above problems,this paper proposes a face image age estimation method based on multi-task learning.Firstly,the background and significance of face image age estimation are introduced,and the deep learning methods for age estimation are comprehensively arranged and introduced;the theoretical basis of the deep convolutional neural network model is introduced;Estimated deep convolutional neural network model structure,detailing the 34-layer residual convolutional neural network model.Secondly,in view of the correlation between ages and the ambiguity of age labels,this paper uses deep label distribution learning to encode ag e labels,uses KL loss function to measure the difference between age label distribution and prediction distribution,and uses Image Net pre-trained network.Age estimation experiments are performed on the MORPH Album2 dataset to improve the accuracy of age estimation.Then,for age estimation,race and gender recognition are related problems.This paper regards age estimation as the main task and race and gender recognition as auxiliary tasks,and designs three multi-task learning models MT-128,MT-256,MT-512,the best structure MT-256 was obtained through comparative experiments on the MORPH Album2 dataset,and then the deep label distribution learning was introduced into the MT-256 model,which further improved the accuracy of age estimation.Finally,for the problem that age estimation requires not only global features,but also local age-sensitive features,this paper designs a multi-scale soft attention module,applies the multi-scale soft attention module to the MT-256 model,and matches the depth label distribution encoding.Combined,by conducting comparative experiments on the MORPH Album2 dataset,the form of the multi-scale soft attention module and the position applied to the model are discussed,and the optimal policy is obtained.Subsequently,experiments were carried out on the Adience dataset using the method in this paper to further verify the effectiveness of th e method in this paper.