A Research Of Facial Age Estimation Algorithm Based On Multi-Neural Networks

With the rapid development of smart devices,facial age estimation technology is of great significance to many aspects such as security monitoring,product recommendation,human-computer interaction,and commercial marketing.However,the aging of the face is a slow change that cannot be detected by the naked eye.Compared with other human biological characteristics,they are more easily disturbed by uncertain factors inside and outside the human body.It is these factors that bring great difficulties and challenges to the study of facial age estimation.Age is a special biological feature.Simply treating the age estimation problem as a classification problem or a regression problem ignores the ordered information of age.And if the age is divided into multiple non-overlapping age groups according to the fixed interval length,the boundary ambiguity between age groups will be caused,which affects the accuracy of age estimation model.Based on the research of existing facial age estimation algorithms,this paper proposes a multi-convolution neural network method and a hierarchical classification network method,hereby realizing two age estimation models.For the ordering problem of age,this paper uses multiple convolutional neural networks to transform the facial age estimation problem into multiple bi-class subproblems.The subproblem is to determine whether the input face image is larger than the preset age.Finally,the results of all subproblems were fused to obtain the age predicted values.Each basic convolutional neural network uses the ordered age label to learn the corresponding age characteristics across the whole training data set.This method expands the training set in a disguised way to prevent model overfitting.All basic convolutional neural network parameters are initialized by a pre-trained convolutional neural network,effectively reducing the time required for model training.Experiments show that the facial age estimation model based on multi-convolution neural network is better than the common facial age estimation model,which shows that considering the age order is helpful to the performance of the age estimation model.At the same time,the age characteristics learned by the basic convolutional neural network can more effectively express facial age information.To solve the boundary ambiguity problem of age groups,a facial age estimation model based on hierarchical classification network is proposed.Dynamically moving or scaling the range of age groups based on the input image can effectively solve the boundary ambiguity problem,while also taking into account the continuity of age.The hierarchical classification network method uses the classification strategy from coarseto-fine to group the face images by age,and each level of the network is a further classification of the upper-level age range.The hierarchical classification strategy makes the neural network model need only a few parameters,reducing the size of the model and greatly shortening the training time.The experimental results show that compared with the complex model,the accuracy of facial age estimation model of hierarchical classification network decreases a little but the memory consumption is very low.Compared with other compact models,the model size is smaller,the precision is significantly improved,and it is more suitable for mobile platforms and embedded platforms.