Face Recognition Algorithm Based On Traditional Subspace Embedding And Deep Learning

Image recognition,as an important field of artificial intelligence,is widely used in modern information society.As an important branch of image recognition,face recognition has great application value and development prospect in the security field as well as in the civil economy and entertainment field due to its natural and concealed characteristics.In recent years,with the continuous development of deep learning technology,deep neural network has achieved great success in the field of image detection and recognition.Image features can be divided into two categories:hand-designed features and automatically extracted features.In this paper,the problems related to face recognition using these two types of features are studied respectively,and two different face recognition algorithms belonging to the studies are proposed.The main work is as follows:(1)For the hand-designed features,based on the theory of random dot product graph,this paper proposes a face recognition algorithm fusing gabor feature,this method is based on the normalized random dot product graph can be a good characterization of the data distribution,using the face image feature extracting from gabor operator as input data of algorithm,then use the random dot product graph to map the input data to low dimensional spherical space,at the same time keep the distance of the image data.Using the supervised category information to introduce the penalty matrix makes the algorithm more robustness and reduces the impact of the isolated image projection on the algorithm recognition rate.The recognition rate of this algorithm on ORL and FERET gray database is better than PCA and LBP methods.(2)For the features automatically extracted by the deep learning network,this paper analyzes the contribution of facial components in face recognition,this paper proposes a deep neural network face recognition algorithm based on segment components.The face image is segmented according to the key regions of the face image.The segmented image is taken as the input data of the deep neural network training model and the feature vector of the segmented image is obtained.Then the feature of the segmented image is spliced and the classifier is trained to recognize.Compared with DeepID algorithm,the recognition rate of LFW database is better,and the network model has the advantages of small scale and good flexibility.