Research On Fine-Grained Age Estimation Method For Face Image

Face images are an extremely rich source of information.As one of the key attribute information of a face,age plays a fundamental role in people's social communication,making the age estimation problem of face images closely related to the actual needs.When classifying face images of different ages in the face image age estimation task,images with similar ages when distinguishing face images have high similarity and slight inter-class differences.Therefore,this paper proposes a method based on competing ratio loss,drawing on the idea of fine-grained image classification.In the image classification task,cross entropy loss only focuses on the posterior probability of the correct class.Hence,it cannot be discriminative against the samples not belong to correct one(wrong classes)directly.So,competing ratio loss is proposed to better discriminate the correct class from the competing wrong ones,which calculates the posterior probability ratio between the correct class and the competing wrong classes to further enlarge the probability difference between the correct and wrong classes.First,the deep convolutional neural network(DCNN)has achieved great success in the face age estimation task.This paper introduces the DCNN network structure and loss function used in the face age estimation task,analyzes the development process of network structure and loss function.The feasibility of designing a loss function to achieve face age estimation based on fine-grained classification is verified.Secondly,this paper analyzes the shortcomings of the cross entropy loss function in in the face age estimation task,and proposes a competition ratio loss function for this shortcoming.It also introduces the definition and advantages of the competition ratio loss function in detail.Finally,competing ratio loss is evaluated in the image classification task and face age estimation task: the face age estimation dataset uses the unconstrained Adience,and the general image classification tasks use the CIFAR,SVHN,and Image Net.The finegrained image classification task uses the CUB200-2011 and Cars196 datasets;Different network parameter settings are made for different datasets;the CIFAR dataset is used to analyze the selection of hyperparameters in the competing ratio loss,and some experiments are performed on networks of different types and depths.Finally,the experiments are performed on general image classification tasks,fine-grained image classification tasks,and face image age estimation tasks.The experiments prove the robustness and effectiveness of the competing ratio loss for different structures and deep networks in the face age estimation task.Competing ratio loss can effectively widen the inter-class difference and improves the accuracy of face age estimation.In addition,competing ratio loss improves the accuracy of other image classification tasks,especially fine-grained image classification tasks.