Research On 2D Face Recognition Based On Neural Network

In the second decade of the 21^st century,artificial intelligence technology has begun to flourish.Among them,face recognition technology has been widely used in many fields such as security,finance,e-government,entertainment and so on,and thus has become a hot research topic in artificial intelligence technology.However,facial images are easily affected by facial expressions,gestures,lighting,and appendages.These factors make facial images extremely variable,increase the difficulty of face recognition,and higher requirements for algorithms accuracy.At the same time,due to the popularity of mobile devices,people not only require high recognition accuracy for face recognition,but also require as much as possible a reduction in running time,which makes face recognition problems difficult and challenging.This thesis separately designs and implements different algorithms for these four steps:face detection based on cascaded ensemble regression trees,face normalization based on affine transformation,face feature extraction based on Inception layer neural network model,and feature match based on SVM classifier.First,a cascaded ensemble regression trees algorithm is proposed in the face image detection.Because on the one hand,reliable features are needed to predict the shape of the face,on the other hand,an accurate face shape estimation is needed when extracting the features.Therefore,the face detection process needs to be carried out iteratively,ie using cascaded ideas.The algorithm cascades several strong regressors,and each strong regressor is composed of several weak regressors.Each strong regressor is an ensemble residual regression trees.Each leaf node of a tree stores a residual regressor.When the input value falls on a node,the residual value is added to.On this input,this will serve the purpose of regression.Finally,when all the residuals are added together,the purpose of face detection is achieved,that is,to return a frame box containing facial landmarks points.Finally,the design experiment proves the effectiveness of the algorithm,and the comparative experiment proves that the algorithm is advanced.Second,the face image is normalized by using affine transformation.The face image is corrected to face the camera and the size of the cropped image is the size of the standard input neural network.Uniformly normalizes the input image to the subsequent feature extractor,reducing unnecessary image noise effects.Thirdly,in the feature extraction stage,a feature extractor based on the Inception layer neural network model was designed using the triplets loss function.The triplets loss function can make the face image of the same person gather in the high-dimensional feature space as much as possible,while keeping the face images of different people as far away as possible.The Inception layer has higher classification accuracy and fewer parameters than the commonly used full-connect layer and convolution layer in the neural network model.After training,the neural network maps the face to a 128-dimensional unit hypersphere space,ie,128-dimensional feature vectors are extracted from each person's face table.A face comparison experiment was designed to show that compared to FaceNet,the leading model in the academic world,the proposed model reduced the accuracy by about 4%,but reduced the number of parameters by about 56%.Fourth,feature matching is based on the SVM classifier.A double face recognition experiment was designed to prove the effectiveness of the proposed algorithm.Then the comparison experiments were designed and compared with the three techniques of Eigenfaces based on PCA dimension reduction,Fisherfaces based on LDA linear discriminant analysis,and LBPH based on local binary pattern histogram.The results show that the accuracy of this paper is higher than the above system.