Human Face Age Estimation Based On Densenet

The age estimation based on face is an important subject in the research topic of face image,and it is also an indispensable research topic in the field of computer vision.Age as an important attribute of face,through age can provide verification information for face recognition,target tracking,identity recognition and other tasks.The face application scene is complex,and human face age estimation faces many difficulties in non-contact age estimation through human face.In recent years,various face age estimation algorithms have been challenging the age estimation task,and the effect is getting better.However,the face age estimation for full-scene application still faces many problems to be solved at present,such as small number of data sets,insufficient feature extraction of age estimation algorithm,insufficient model optimization,etc.In this paper,the following innovations are proposed for the shortcomings of the previous age estimation algorithm:(1)In this paper,we extract the age characteristics of faces using DenseNet,and finally get the face age estimation model with high accurate rate and loss reduction better than other neural network algorithms.The age estimation model of DenseNet has few calculation parameters and high feature utilization rate,but it has the disadvantages of high fitting degree,comparing with the current mainstream algorithms.Based on this issue,the DenseNet-ED(Exponential Dropout)is proposed for the dropout layer drop factor p exponential function.The improved post-DenseNet reasonably discards the characteristics of different network layer extraction.The feature drop rate of dense block extraction increases with the number of layers increasing,which increasing the proportion of basic features and reducing the over-fitting of the model.The model estimates the face age is more accurate after the extracted features are mapped.(2)In order to increase the model accuracy and the degree of loss reduction of DenseNet,this paper proposes a learning rate oscillation shrinkage tuning algorithm to train and tune the DenseNet-ED model.The tuning algorithm can make the network avoid falling into the local minimum in the process of learning the optimal solution of parameters during training,make the network parameter learning close to the optimalsolution of parameters,and improves the robustness and generalization ability of the age estimation of DenseNet-ED model.This paper has a good effect of 70.013% accuracy and 4.9813 average absolute error in face age estimation on IMDB-Wiki data set,and the age estimation effect of face estimation is the best by using the age estimation model of this DenseNet compared with the other neural network algorithm models.