Methods Of Feature Extraction Based On Manifold Learning In Face Recognition

Face recognition is one of the most potential research subjects in the field of biometricidentification technology. Face recognition involves much knowledge of other disciplines, andhas broad application prospects in national security, military security and economic field.Therefore, the study of face recognition technology is great theoretical and practicalsignificance.Extract effective discriminant features is a key factor for the classifier in face recognition,which requires reduction of the data dimensionat and keep the face data set of the originalnature of structural characteristics unchanging at the same time. In this paper, based on theinveatigation of feature extraction method based on manifold learning, the main contents andinnovations are listed as follows:Firstly, aiming at the drawback that Local Fisher Discriminant Analysis (LFDA)algorithm and the Marginal Fisher Analysis (MFA) algorithm solve the problem ofmultimodal data classification. An adaptive algorithm to select a close neighbor of the point isutiliy in this paper, so it can more accurately select the same class neighbor points anddifferent class neighbor points, and perfectly detect the intrinsic geometric structure ofmanifold.Secondly, aiming at the Marginal Fisher Analysis (MFA) algorithm is inadequate when itkeep the intrinsic geometric structure of images, and it is aUnsupervised Discriminant Analysis Algorithm, which only utilizes labeled data and wastesrich unlabeled data. Therefore, this paper combined the MFA algorithm and Tensor to improveto be Semi-supervised Discriminant Analysis Algorithm Based on Tensor. The method adoptsthe two-dimensional tensor to show the image samples, so it can perfectly detect the structureof data manifold. Moreover, it sufficiently utilizes the labeled data which containsdiscriminant information and rich unlabeled data, which make the same class points class tobe closer, and the different class points is distant in the low-dimensional space.Finally, the Unsupervised Discriminant Projection (UDP) algorithm is sensitive to outlierand noise, so it is inadequate when it keep the intrinsic geometric structure of images.Besidesthis algorithm has small sample size problem.So, this method, combined the UDP algorithmand the maximum scatter difference criterion, introduce into robust path similarity to measure similarity of data. So it improves the algorithm robustness to noise, avoid the small samplesize problem, and enhance the stability of the algorithm.