Research On Kernel-based Discriminant Analysis Algorithm

Face recognition technology is used in computer analysis for face image. It can extract identification information to recognize people. Face recognition technology relates to many research fields, such as technology image processing and analysis, computer vision, artificial intelligence, pattern recognition, biology, so it has a very broad application prospects. Because of the particularity of face image, face recognition technology remains many unsolved problems.Face recognition includes image preprocessing, feature extraction and recognition, while feature extraction is one of the most basic researches in face recognition. For image recognition, extracting effective imaging features is the first important problem to solve. Nearly, kernel feature extraction methods become very popular in feature extraction. This paper proposes a fast kernel feature extraction method based on virtual samples.Kernel feature extraction methods aims at taking the original space linear inseparable problem into a higher-dimensional space linear separable problem, however, the projection vector is usually spread linearly by all the training sample in most cases, so it spends a lot of time to calculate the huge kernel matrix, which makes kernel feature extraction methods rather time consuming. In order to solve this problem, some kernel accelerate algorithms have been proposed, but in most cases, searching projection vectors in these accelerate algorithms is also very time consuming. In order to reduce spread elements, these speedup method using iterative algorithm from the original sample to choose the elements. That is a very expensive process, especially considering the calculation of each kernel function, it becomes even greater. Furthermore, because of the abundances of the sample information, the recognition rates of these accelerate kernel methods all declined.This paper put forward a new idea to accelerate kernel method: kernel feature extraction approach based on virtual sample. This approach calculates a smaller kernel matrix by constructing a few virtual samples in input space and expresses kernel discriminant vectors using three types of MVS, which include mapped eigen-samples (MES), mapped mean samples (MMS) and mapped common vector samples (MCS). Experimental results on the public AR, FERET and CAS-PEAL face databases show that this approach is effective in both saving computing time and acquiring favorable recognition results.