A Comparative Study Of Face Recognition Algorithms Based On Small Sample Data Sets

Computer and information technology are rapidly integrating into people's daily life.Due to the unique characteristics of biometrics,it promotes the development of new recognition methods,which has attracted people's extensive attention and has been applied in security payment,unlocking,case detection and other fields.Although face recognition technology has achieved great success,it still faces some challenges,such as face similarity,pose,illumination,expression,occlusion,age change,change of imaging scene and ethnic face feature recognition.The main challenge comes from the collection of face image data.In the process of small data collection,the human cost is less.However,most of the current face recognition algorithm models need a large number of sample data to perform best,but it is difficult to collect such data sets in reality.Some mainstream face recognition algorithms are difficult to be widely used in real scenes.First of all,the sample data in the training data is limited in the number of samples for model training,which leads to over fitting phenomenon.Secondly,when the sample image has large similarity and complex structural features,It is possible that the class distance between samples is less than the in-class sample distance,which leads to the false recognition or the recognition accuracy is not obvious.Another point is the problem of algorithm bias,that is,different algorithms in the same kind of face recognition accuracy is different,each has its own advantages and disadvantages.Therefore,it is a very challenging research problem to analyze the recognition performance of different algorithms,how to choose the best recognition method,how to extract effective discriminant features from a small number of diverse training samples with large similarity of face image,and how to ensure the real-time performance of face recognition model and the robustness of the algorithm.At the same time,it plays a more effective,safe and reliable role in practical application.In view of the above problems,this paper makes the following research:1)Considering the current lack of publicly available and diverse ethnic face data sets,In this paper,the small sample ethnic face data set composed of face images of six different ethnic students in our university is further preprocessed and improved.2)In order to express the theme of this study more clearly,the paper first expounds the traditional subspace methods,such as PCA,LDA,LPP and their improved algorithms 2DLDA?2DPCA?2DLPP?RM2DLDA?RM2DLPP,etc.Then introduces BP neural network and PCA+ BP,The basic theoretical knowledge of PCA + BP + KL,and other hybrid algorithms is introduced.Finally,the classical lenet-5 convolutional neural network is listed to introduce the theoretical knowledge based on convolutional neural network.3)The cosine similarity measurement method is used to analyze the similarity between the face images of the same ethnic group.A large number of experiments are carried out to verify that the similarity of the face image of the same ethnic group in the face image data set with ethnic tags is significantly higher than that in the face image data set without ethnic tags.Finally,the conclusion is drawn through the comparison experiment: Because of the higher similarity between the face images of the same ethnic group,the task of facerecognition is more difficult.Finally,an improved face recognition algorithm of RM2DLDA+cos is proposed,which can effectively improve the face recognition accuracy of small sample ethnic face recognition.4)A large number of experiments were conducted to compare the trained PCA+BP?PCA+BP+KL and other mixed algorithm models and the retrained models adapted to the small sample ethnic face data set,and it was found that the PCA+BP+KL algorithm was most suitable for the small sample ethnic face data set.5)On the basis of lenet-5 neural network,a simple convolution neural network model adapted to the small sample ethnic face image data set is designed by adjusting the network parameters through a large number of experiments.The network is composed of two convolution layers,two pooling layers,one full connection layer,one softmax output layer,The method of dropout effectively avoids the problem of overfitting,and the size of the convolution kernel is 5×5,The number of convolution kernels of the two convolution layers is10,the average pooling replaces the original maximum pooling,each excitation layer uses Relu function instead of sigmoid function,reducing the number of full connection layers,The number of neurons in the full connection layer was 170.Finally,Softmax classification function is used for classification,which can automatically extract face features and carry out recognition,It effectively improves the accuracy of small sample ethnic face data set.6)All the experiments were done with widows10 as the test system combined with matlab and python under the tensorflow framework.In order to solve the problem that the computation speed of convolutional neural network model is too slow in the training process,GPU(GTX1050 graphics card)is used instead of CPU in the training of convolutional neural network.