Deep Convolutional Neural Network For Face Recognition And Face Analysis

Face is one of the most direct representations and most natural biological features of people.Face recognition is a biometric technology that uses human facial information to identify them.With the development of science and technology,face recognition has been applied to many aspects,which has brought great convenience to our life.ID face recognition is a specific application of face recognition,which is to identify people by comparing with ID photos.It is similar to the normal face recognition,but with slight difference.Face recognition often accompanies by face analysis which is to predict gender,age,expression,etc.by algorithms and can help for further judgement and application in practical scenes.In recent years,deep convolutional neural networks have been widely used in computer vision,greatly improving the performance of classification problems.Researchers mainly focus on how to build deeper models,get more data and design better loss functions,so as to get more expressive deep features.We usually use softmax loss function to supervise the training of convolution neural network,and then use the output of the last layer before softmax loss as features.In order to learn more discriminative facial features,this paper focuses on the design of network structure,optimization of loss function and multitask learning.The main contents are as follows.First,we analyze the softmax loss function,and improve it mainly by manipulating the cosine value and feature norm.These methods do not change the principle of the softmax loss,so can still be optimized by typical stochastic gradient descent.We use the handwritten digits recognition as example to visualize the features,and study the effects of different parameters.Then we use the CASIA-WebFace database to train the face classifier,with a 28-layer convolutional neural network refer to popular and advanced architectures.The improved softmax loss is applied to thetraining process and the model is tested on the LFW.Results show that our improved method is better than the original softmax loss function,and can get good accuracy on LFW with one single model.Next,we further research the ID face recognition because the model trained above can not get a good performance on the ID face problem.We collect a large number of ID face image,in which most people only have one ID face photo and one normal photo.Because the ID face often come in pairs,we modify the classical metric learning method Contrastive loss and Triplet loss to train the network.We also propose a method,named Fisher loss,based on the distribution of positive and negative pairs,considering the characteristic of the ID face.It pushes the distribution of the positive pairs away from that of the negative pairs,thus improves the discrimination of features.Results show that our method can effectively improve the accuracy on ID face test set.Finally,we use the multitask learning framework on face analysis,combining face recognition with gender recognition,age recognition and expression recognition.By sharing the previous convolution layer,the network can learn more general filters.In addition,training different kinds of databases at the same time can reduce the overfitting problem of the network.Results show that multitask learning can improve the performance of face analysis task,while individual task training ignore the connections between different tasks.