Locally Linear Embedding Based On Manifold Learning And The Applications In Face Recognition

Face recognition is a biometric identification technology, because of its ease of collection of data, facial intuitiveness, as well as strong representation, it has now become the pattern recognition and artificial intelligence research hotspot.But the face during processing often encounter this problem:First, face images vulnerable to external factors such as light, facial expressions, gestures and other effects, last the feature extraction for human face recognition will encounter greater difficulties, due to the face image is a high-dimensional data.Research has shown that human face in a sense is a nonlinear manifold structure, which is mainly affected by some internal variables control, if the intrinsic variables about illumination, facial expressions, gestures, etc can be found from the face manifold, it can effectively for high-dimensional face data dimensionality reduction.Manifold learning is a nonlinear dimensionality reduction method, it is a good way to keep some of the characteristics of high-dimensional data, and the goal is to find the low-dimensional manifolds embedded in the high-dimensional space.In this paper, a manifold learning typical method-Locally Linear Embedding (Locally Linear Embedding, LLE) has been improved, and it combined with a kernel method KFDA (Kernel Fisher Discriminant Analysis, KFDA) to apply to the face identification, and achieved good results.The main research work is as follows:1.In this paper, base on the image processing, manifold learning similarity measure LLE algorithm has made improvements.Since Euclidean distance can not be measured in image space two sample points real distance, we propose to geodesic distance and image euclidean distance instead of the traditional euclidean distance.2. For the first step of the original LLE algorithm Neighborhood number k value problem, we have made improvements. LLE algorithm assumes that all of the sample is uniform, so all the sample points are taken the same number of neighbors. But it is difficult to meet the above assumptions in practice, most of the sample data are non-uniformly distributed, the face image as well. To solve this problem, this paper proposes an adaptive local linear embedding algorithm, based on each sample around points distribution, respectively, to find the most suitable number of neighbors.3. After using the improved LLE algorithm for face feature extraction, non-linear discriminant method KFDA is used as a classifier to classify the human face, that is to project low-dimensional space of non-linear data into a high dimensional space to achieve linear separability, owing to useing of kernel methods, the calculate complexity does not increase compared with the linear the method, and it has a better effect.