Research Of Face Recognition Method Based On Few Sample

With the rapid development and application of artificial intelligence technology,many fields continue to promote intelligent changes drive by this technology.Biometric recognition,as a technology that uses individual physiology or behavior for precise identity matching,has once again become a research hotspot with the promotion of artificial intelligence technology.Among the many biometrics,the features that uniquely match the face and individual identity occupy an important position in the field of biometric recognition.However,the problem of face recognition with limited samples is currently a difficult problem,because most excellent face recognition algorithms are datadriven,and collecting large amounts of face data is a difficult task.Sometimes,this directly affects the performance of face recognition algorithms.Therefore,how to solve the interference caused by limited samples on face recognition will be the focus of this paper.Under the condition of limited training samples,this paper conducts research on sample enhancement and discriminative face feature extraction algorithms,with the aim of improving the accuracy of face recognition under limited samples.The main work of this paper is summarized as follows:First,the basic framework of face recognition algorithm with limited samples is introduced,and the key technologies involved in each link in the framework are compared and analyzed.Secondly,a deep face feature extraction algorithm based on multi-task convolutional neural network is proposed.This method is also based on transfer learning.In order to further extract discriminative face features,thereby solving the problem of face recognition rate under limited samples,a multi-task convolutional neural network is introduced,that is,by adding face key point positioning task.The face key point positioning task is added to assist the face recognition task,and more discriminative face features are extracted.Then,the face classification task is combined to form a multi-task convolutional neural network model that includes classification and regression.Finally,we proposed a limited sample face recognition algorithm based on transfer learning and feature augmentation.Since learning a robust feature extraction model requires a large amount of training data,transfer learning is introduced in this paper.First,by training a deep convolution neural network on a multi-sample common data set,and then using the trained model to extract target dataset of face features;secondly,using feature expansion algorithms to increase intra-class differences in limited training samples;finally,using the expanded feature samples to train a classifier to achieve face recognition.