Based On Kernel Method Of Nonlinear Dimension Reduction Method Of Face Recognition Research

Face image data are usually high-dimensional. High computational complexity and high storage capacity brought by high-dimensional data increase the difficulties of identification. However, the intrinsic dimension of the image space is generally low, which means high-dimensional image data can be expressed by low-dimensional intrinsic data. Therefore, the extraction of effective image low-dimensional intrinsic data is the key task to accomplish image recognition. Commonly used linear dimensionality reduction methods, such as Principal Component Analysis and Linear discriminant analysis, can not discover the non-linear structure of the original high-dimensional data sets and the intrinsic characteristics of image data. The limit of linear dimensionality reduction methods affects the accuracy of feature extraction, and leads to a low recognition rate. However, nonlinear dimensionality reduction method can discover nonlinear structure hidden in high-dimensional data set and maintain the intrinsic structure of the data. The accurate feature extraction of nonlinear dimensionality reduction method can improve recognition performance.Dimensionality reduction based on kernel methods is a commonly used non-linear dimensionality reduction method. It not only has the advantages of nonlinear methods, but also has a more clear and concise form.By summarizing the commonly used linear dimensionality reduction method and nonlinear dimensionality reduction method based on kernel methods, this article improves the correlation algorithm and obtains the following findings.1. This thesis brought up the learning algorithm based on adaptive dual fusion. Dual-core refers to the Gaussian kernel and polynomial kernel. The Gaussian kernel can prominently reflect the information of the adjacent samples, while the polynomial kernel can comprehensively reflect the information of all samples. The adaptive fusion of the former and the latter not only can overcome the deficiency of H.L.Xiong’s methods, which is based on single kernel for optimizing, but also can give full play to the advantages of Gaussian kernel and polynomial kernel.2. This thesis brought up a face recognition algorithm based on a kernel generated by the training sample information. This method can generate a kernel function through some certain ways by any given function. It can also optimize the generated kernel function in the empirical feature space. Then we do feature extraction on the high-dimensional image data based on kernel methods. This method can overcome the deficiency of the usual face recognition based on kernel methods which must be established on the basis of the kernel function to do feature extraction, at the same time, it also make full use of training samples.