The Application Of Graph Embedding Method In Face Recognition

Since entering the21st century, the momentum of development of informationtechnology is more and more rapidly, the development of computer technology is movingtoward intelligent and automatic, and the application of video surveillance technology hasbecame more and more widely, and the automatic identification technology under videosurveillance is needed to achieve non-contact, remote, hidden identity authenticationtechnology that can quickly identify the identity of the person in the scene,which can get ridof the traditional methods’constraints that invest a lot of manpower and resources in the past.Thus, regardless of efficiency, or investment, human face identification technology will be thebest solution under video surveillance identification technology. Extract real-time monitoringscreen from the video surveillance, use face detection technology to find faces, and then usethe database of human face images stored do comparison, which can identify identifiableinformation quickly and easily. In recent years, face recognition technology has begun toemerge in the intelligent automation technology, get the attention of the relevant researchers,and occupies a very important position in the field of image processing and artificialintelligence.Face recognition deal with human face picture, a picture of resolution of3232canrepresent data of1024pixels, we use three integer of0255representing each pixel in RGBcolor space, for example,(0,255,0) expresses green. These data are stored in the computerby hardware and must be loaded into memory when they participate in calculation. Assumingthat there are N face images, then the total number of dimensions is1024. When isenough large and the resolution is high, it will cause the curse of dimensionality. Thus, thegeneral procedure to deal with such issues is transforming the high-dimensional data space toa low-dimensional data space (subspace) through a projection transformation and then we canhandle the classification problem in the new low-dimensional subspace, which can reduce thecomputational complexity greatly, and if the obtained feature subspace is good enough, theclassification would be better. Numerous studies show that: Any data space high-dimensionaldata space has potential low-dimensional structural information, through dimensionalityreduction, we can more accurately extract the facial features, but also greatly reduces the computational complexity, and data dimensionality reduction problem is important anddifficult to solve for face recognition. There are many existing data dimensionality reductionmethods, one of them based on graph due to the simple, intuitive, efficient, etc., more andmore researchers are beginning to engage in dimensionality reduction method based on graph.Graph-based algorithm can be unified into a graph embedding framework, used only whenneeded to change their expanded form, such as linearization, tensorization, kernelization andso on.This paper describes the graph embedding method in face recognition. Graph embeddingmethod treat data as nodes in the graph, each node represents one of the sample data, use aundirected-weighted graph to describe the relationship between two nodes, through assigninga weight between two nodes to represent a neighbor-relations. Using a graph to represent therelationship between data, that is the data space is seen as a "manifold" point in space, firstlywe assume that these data points are in a high-dimension "manifold" space, and then use theneighbor relations of nodes in graph to find a reasonable description method, or an objectivefunction, to find a graph of low-dimensional space to approximately represent the originalspace, and data after dimensionality reduction can be reserved. Good or bad graph modelcould effect the result after dimensionality reduction, so we are committed to construct areasonably and accurate graph to describe the relationship between nodes. Generally, we usean adjacency matrix to represent a graph. In this paper, in the framework of graph embedding,we try to improve the construct method of graph, when calculating whether the nodes are"close",we use two-dimensional face image as a input, the number of columns(or rows)which are neighborhood as a parameter r, to determine whether these two images are or not"close". Cleverly use of the local features of face images to compensate for the lack of globalfeatures, which can make the weight matrix G could accurate describe intrinsically relationsaccurately. Through a large of experiments on the YALE, ORL, AR, etc. show that ourmethods improves the accuracy of the weight matrix; in LPP, NPE’s test results show that ourmethod is very effective to improve the recognition rate.