Research On Face Recognition With Single Sample Per Person Using Joint Local And Global Depth Features

In recent years,with the rapid development of society and technology,face recognition has been implemented in many aspects of our life.Most of the existing face recognition systems are based on multi-sample training,in this case,each person has enough samples,which provide more information for training to deal with different face changes.However,in some special scenarios,such as law enforcement,passport verification,identity card verification,etc.,each person can only get one training image,so the face system can only get limited information during training,which often fails to complete subsequent recognition tasks.Aiming at the situation that there are limited number of images to train the face system,face recognition with single sample per person(SSPP)is derived,which usually refers to that each person only stores one face image as a training set to recognize the identity of the test image with changes in pose,illu-mination and expression.At the present stage,SSPP still faces many problems and challenges.On the one hand,traditional methods based on single sample per person usually ignore the exploration of more discriminative features,and the method based on deep learning has not been well studied for the single sample per person problem.On the other hand,the information of the gallery set is usually ignored and only be used for recognition stage in most cases.In order to better solve these problems,this paper focuses on two aspects of the research on this topic based on current research situation.Firstly,in order to obtain more discriminative features at the stage of face feature extraction,we propose an adaptive convolution local and global learning network based on deep learning.On the basis of the original deep learning network,the proposed network deeply mines the local in-formation of human face by dividing the feature maps into dense samples inside the network,and combines all the local features to generate the global feature,thus we extract regional adaptive convolution features which are locally and globally discriminative to face identity and robust to face variation.Then,a novel class-level joint representation framework is presented to exploit the distinctiveness and class-level commonality of different facial features,and complete the classification finally,in which both discriminative facial features robust to various facial variations and powerful representation for classification with generic facial variations have been fully exploited.By contrast with some advanced methods with single sample per person,the exper-imental results show that the proposed framework can obtain more representational features,and the higher recognition accuracy.On the other hand,in the face recognition problem with single sample per person,for the purpose of adding the gallery set's information into training stage,which further improve the adaptability of features,we attempt to add the prior information of the gallery set by two methods.First we using the similarity of different face changes,the gallery set was amplified based on the face changes from the sample in the additional training set which most similar to the every each gallery.Moreover,we use the generative adversarial network(GAN)which shine in recent years,and try to maintain sample identity information.In addition,according to the similarity of face changes between different identity,we propose a weight-embedded supervision in training phase for finely learning adaptive deep feature for the single-sample gallery set,which adaptively learn the deep features by introducing the information of thegallery set,and use the attention mechanism to redistribute the weight of partitioned regions.The experimental results show that the proposed method of introducing information by using sample amplification can improve the performance to some extent,while the proposed method of introducing information of gallery set by weight-embedding implicitly shows good performance,especially in more complex and difficult databases.In short,this thesis analyzes the face recognition with single sample per person,and proposes an adaptive convolution local and global learning network based on deep learning and class-level joint representation in the feature extraction and recognition stage respectively,.Then an adaption of facial features to the single-sample gallery set is conducted by our designed weight embedding.The proposed method has been evaluated on several popular databases,Experimental results demonstrate the much higher robustness and effectiveness of our methods compared to the state-of-the-art methods.