A Research On Age Estimation Algorithm Based On Machine Learning

With the rapid development of Internet and artificial intelligence,people pay more and more attention to customization and personalization of services.Age,as an important biological feature to judge personal hobbies and habits,makes age estimation a hot research topic.However,the face image is affected by many factors such as light,posture and so on,the difficulty of age estimation is deepened.Therefore,age estimation based on face recognition is a very valuable topic.With the rapid development of computer vision,age estimation has a new idea.Convolutional neural network has become an important direction of age estimation.In this paper,based on convolutional neural network and VGG-16(Visual Geometry Group)network,a coupled evolutionary network based on DEX(Deep Expectation of apparent age from a single image)is proposed to estimate age from two aspects: label distribution and regression expectation.The specific work is as follows:1.The methods and steps of face image preprocessing are given.For face detection,the Adaboost method based Haar-like feature is used to recognize the face area in the image;for image standardization,affine transformation is used to correct the direction,and the context information is increased by edge expansion.Finally,face image preprocessing is carried out using OpenCV library.2.An age feature extraction model based on VGG network model is presented.By analyzing the structure and performance of each VGG network model,this paper determines the network model VGG-16 for age feature extraction,and describes the process of age feature extraction.In view of the inefficiency of VGG-16 network training,dropout method and fine-tunning method are used to improve the training efficiency.At the same time,5 epochs are trained.The time of network training is shortened from 39562 s to 27245 s,and the error of age estimation is reduced from 10.862 to 5.562.3.A coupled evolutionary network model is proposed for age estimation.Based on DEX which is a classical model of age estimation,a coupled evolutionary network is proposed in this paper.On the one hand,the last output vector of the network is regarded as the age label distribution,and the KL divergence based on the previous label distribution is used to evolve the age label distribution learning.On the other hand,the relaxation regression is carried out according to the regression expectation and the loss factor of the previous one,which transforms each training target into an age label range rather than an accurate age label value.Compared with DEX model,the age estimation error of coupled evolutionary network on Morph database is smaller.4.The superiority of the model is verified by setting up comparative experiments.Based on tensorflow framework,parameter selection experiment,age feature extraction experiment,fusion model comparison experiment and coupled evolutionary network model experiment are set up.The best performance parameters are selected through experiments,and compared with other classical algorithms on Morph and MegaAge-Asian datasets,it is proved that the coupled evolutionary model has higher accuracy of age estimation.