Face Recognition With AMSoftmax Loss And Center Loss

For image recognition tasks,a popular model is to combine convolutional neural network and softmax cross entropy loss.However,the “overconfidence” nature of the softmax function can lead over-fitting and misclassification problems to the model.Face recognition needs to calculate the distance between the features and determine the similarity for identification.Based on that,this thesis analyzes the softmax cross entropy loss for improvement.Firstly,we construct a simple convolutional neural network structure and train the network on MINIST dataset,and get the two-dimensional features map.It is proved that the softmax cross entropy loss function does not apply to the task of face recognition by analyzed the cosine distance and euclidean distance of the features.In order to explore the effectiveness of increasing the distance between classes on the face recognition task,this thesis makes the same experiment by the AMSoftmax cross entropy loss function.The result shows that only improve the separation between classes not make the face recognition perform better.Finally,this thesis proposes a joint AMSoftmax with CenterLoss called AMS-CL which considers not only inter-class separation but also intra-class compact.The AMS-CL is work by analyzed the features map of MNIST dataset.This thesis constructs the AMS-CL model by joint the improved ResNet50 network model and AMS-CL cross entropy loss.Training the model with FaceScrub dataset,test the result on the LFW dataset which have been deduplicate data with training set.The experiment show that the accuracy of face recognition by AMS-CL model is more than Softmax model,CenterLoss model and AMSoftmax model.Increasing the distance of inter-class and reducing the intra-class distance can effectively improve the accuracy of face recognition.At the same time,in order to explore the influence of the central loss constraint on the overall model,this thesis control different central loss coefficients to do the experiment,and the results show that the model works best when the coefficient is set to 0.01.At the end of this thesis,the face recognition system which based on the AMS-CL model has been designed and implemented,it is developed by the Python and QT.This system has many modules such as user management module,video stream module,face detection module,face recognition module and face information management module.It supports the entry and recognition of new face attribute information.At the same time,the system can run smoothly in the GPU-free environment,and the actual test also has better performance.