Graph Regularized Non-negative Matrix Factorization Algorithm Applying To Face Recognition Technology

In the field of biometrics, face recognition technology occupy a pivotal position. Face recognition technology has been widely applied to various fields, which make it rapidly increase the portion of the field of biometrics, more and more close to our daily lives. But we also know that the development of face recognition technology is the process of gradually from the ideal environment to complex environments，the dispose of light, blocking and other issues bring challenge of the accuracy of face recognition. domestic and various algorithms are proposed are designed to address these issues.The algorithm proposed is the optimization of face recognition algorithms in a complex environment. We know that the external environment can not be as ideal as the most traditional "attendance", and this stage is not a simple application of technology. Security, monitor, bank, insurance and other areas all use this technology, as we know Changchun "304" baby missing case, we monitor the suspect in a complex environment, and angle occlusion problems gave recognition technology challenges, so we hope to solve these problems through the optimization algorithms step by step.This article improve and optimize the popular non-negative matrix factorization algorithm, All elements in the matrix decomposes in a non-negative condition. The result is a sparse representation, so it illustrates Non-negative matrix factorization algorithm can better handle occlusion lighting and other issues. Since this algorithm is equivalent to accumulate each module of the face to form human face images. Such as we observe everyone else that many times we will first notice the more prominent feature, which is equivalent to remember a person’s appearance starting point, and then to have a comprehensive understanding of his other aspects. Because of this perception habits of human beings, so we say that the advantage of non-negative matrix factorization algorithm is similar to the perception habits of eyes. Through the introduction of its manifold ideas, it not only to ensure the benefits of non-negative matrix algorithms, but also to ensure the raw data still has its original position relationship in the subspace after dimensionality reduction. Manifold thought can not only be introduced in manifold non-negative matrix factorization algorithm, but also integrate the more popular gradient descent algorithm. This article will integrate non-negative matrix factorization algorithms decomposition algorithm which is based on manifold with gradient descent algorithm, it will improve the recognition rate of the algorithm.This article present a kind of face recognition algorithm which fuse adaptive step gradient descent with manifolds of non-negative matrix factorization through fusing them. we know that the traditional gradient descent algorithm very depend on the choice of the step size, inappropriate steps will make the algorithm converges to a very bad result, Therefore, we introduce a variable step size of the gradient descent algorithm to avoid the influence of algorithm for recognition rate from the selection of the step size. The algorithm can also retain the adjacency of sample points in original space as the traditional manifold NMF algorithm,but also cross less extreme point by using step adaptive algorithm. The variable step size in the iteration gradient descent algorithm can span a poor local minimum value, and converges to local minima better results. Through experiments on AR, ORL, GT database, experiments show that the proposed recognition rate of algorithm is higher than that of the comparing algorithms in repeated run average experiments.