Research And Application Of Age Classificaton By Face Based On Deep Learning

Gender identification and age classification have good application prospects in the fields of identity search,recommendation system and population statistics.With the continuous development and breakthrough of techniques about data collection and comprehension as well as deep learning,there is a great research space to improve the performance of methods related to face recognition,especially the accuracy of age estimation.The purpose of this article is to design a good classification model to judge its age based on someone's face.Firstly,using the fully convolutional layer to solve the problem that the original fully connected layer can only accept the image with fixed size as input.Then,designing and training a neural network of 18 layers with residual modules and fully convolutional layers,which is used for gender identification.The performance of this network is compared with traditional network without residual learning,which proves the validity of the residual module to improve the performance of neural network.In order to verify the introduction of the residual module can effectively alleviate the degradation problem,the residual network and its corresponding network that removes the residual learning are both deepened to 34 layers,and the the results of accuracy are compared to verify the effectiveness of residual learning to solve the problem of degeneration.In order to design a more accurate model of age classification,the residual model above is improved from two perspectives.On the one hand,from the perspective of network depth,by improving the structure of the residual module,it can further overcome the degradation caused by the problem of the network continue increasing,and then we introduce a bottleneck structure.Bottleneck structure can increase the depth of the network structure as well as model accuracy without increasing the model parameters and the amount of calculation.Then,from the perspective of neural network's model branches,the single-channel bottleneck residual module is splitted into multiple parallel branches of the same structure.Under the premise of the same amount of computation,training the age classification model and comparing its performance with the original unbranched network.It is proved effective in improving network performance by increasing the number of network branches.The paper also proposes the possibility that gender may influence the accuracy of age classification.The experimental result shows that the accuracy of the models that training face images of both men and women respectively is 1.1%and 1.3%higher than the model that using dataset of mixed men and women,which confirms the visual age difference between men and women.Finally,the paper designed and implemented a classification system of facial age based on gender,and respectively implemented the face detection module,gender recognition module and age classification module.The experimental results show that the accuracy of the classification model that is improved is enhanced by nearly 6%compared with the traditional residual network model,and the CNN's limit of using images of fixed size as input is eliminated,which has strong application value.