Kernel Function Parameter Research Based On Face Recognition Using KPCA

Kernel principal component analysis is a widespread algorithm on face recognition. During the process of recognition,there are many researches using the kernel function and the kernel method,but how doesn't have a good method to be able until now the effective instruction to choose the most superior kernel function and the parameter with the effective instruction until now.At present researchers mostly adopts experimental means methods to analyze and solve the problem in the specific application area.The choice of kernel function and the parameter still had considerred to be a challenging question.This article carries on the present situation of face recgnition using the KPCA, mainly from the kernel function and,the parameter,the training sample,and the sorter choice to conducts the analysis research,by obtainings the most superior solution and the best parameter sector interval,avoidings in doing before the recognition experiment to do the excessively many parameter selection work before the recognition experiment.1.The choice on kernel functionThis article uses three kinds of kernel function structure with different kernel transformation matrixs,and discovers that when the election took the multinomial nuclear function and the neural network nuclear function,were extremely harsh for the parameter choice,extremely easy which easily leads to create the morbid state nuclear matrix and produced the negative characteristic root and the negative characteristic vector that fail to construct the KPCA transformation model,caused KPCA to transform the model construction defeat;comparatively speaking,the radial direction base on kernel function transformation matrix positive definiteness is extremely good,and the adaptation parameter can be widely applied.So the nuclear function is selected as the transformation matrix in this article.2.Experiment of choosing the paramete of kernel functionWe carry on the experiment of relations between the parameter which selects radial direction base kernel function and the recognition rate.Here,we implement the experiment as follows:the ORL face database is made up of 400 images of 40 individuals and 10 images of each person,we randomly selected five images for training and other five images for test per person(altogether 252 combination as the test sample,we select 10%about the sample,quantity of sample take 30 groups),for the better definite radial direction base nuclear function parameter and the recognition rate relations,we takes[0,20001]this big interval to test,finally tested has obtained the good result.3.The choice of the sorterThe choice of the sorter on face recgnition,the people usually choose Euclidean space distance classifier to carry on the experiment from the sorter,this article use of the Euclidean distance classifier and cosine distance classifier which carry on the experiment from the sorter,obtains two groups of different recognitions rate,which shows that the recognition rate and the classifer choice is concern.