Facial Age Estimation Method Based On Convolutional Neural Network

With the rapid development and improvement of deep learning and related technologies,computer vision will be more widely used.Among them,facial age estimation is a very important branch of computer vision,which has great potential application value in public security,human-computer interaction,information collection,video retrieval and image screening,and has very important research significance.In this paper,by analyzing the existing face age estimation methods,we study from three aspects:the age data set,the age marking method and the network structure,in order to improve the generalization ability of the facial age estimation model and reduce the model parameters.The main research work is as follows:(1)To solve the problem of small number of facial age data sets and unbalanced data,the data sets are enhanced.In order to maximize the performance of the mining model and enhance the generalization of the model,firstly,this paper enhances the existing data set to increase the number and diversity of the data set,and then collects 31876 Asian face images from the unrestricted scene,which are used as the later fine-tuning of the model to make up for the lack of Asians in the data set.Experiments show that the prediction accuracy of face age in Asia is greatly improved by using the data set enhanced model.(2)The method of segmented normal distribution based on age range is used to mark facial images.According to the difference of aging speed and aging mode of different age groups,the normal distribution of adjacent age weighting is carried out for the divided age groups.To solve the problem that the traditional marker ignores the relationship between adjacent ages and the different aging speed and way of face in different stages.The experimental results show that the average absolute error of the model trained by this method is 0.45 lower than that of the model not used.(3)Based on convolutional neural network,a lightweight facial age estimation network model is proposed.In order to improve the efficiency of the model and reduce the amount of computation,this paper uses the improved bottleneck structure and deep separable convolution layer to replace the standard convolution layer.In order to improve the accuracy of neural network model,this paper embeds the residual network structure into the improved bottleneck structure to solve the problem of deep neural network degradation.In addition,this paper improves the compressed excitation network structure,replaces the full link layer with the convolution layer,and selects the appropriate compression ratio according to the number of channels in the layer,and then integrates the improved compressed excitation structure into the improved bottleneck structure.The network can learn the contribution degree of each feature channel in the whole process of feature map extraction,and improve the ability of feature expression.The experimental results show that the network structure proposed in this paper has better performance in prediction accuracy and model parameters.