Design Of Face Recognition Classifier Based On Regression Algorithm

Face recognition technology has been applied to some special areas of our daily life in nowadays,but is not popular in mobile terminal authentication and payment applications.There are two reasons here: Firstly,the accuracy needs to be improved under different light,posture,expression conditions.Secondly,the computational complexity of proposed method which has good effect on the face classification under complicated environment is too large.So many proposed classifiers may not be fit to be used in small memory mobile phone.The purpose of this paper is to improve the recognition rate under different illumination,angle and expression condition by optimizing the classifiers with little computation complexity.In the field of pattern recognition,based on the Nearest Neighbor classifier,the most famous two are the Nearest Feature Plane(Nearest Feature Plane,NFP)and the Linear Regression Classifier(Linear Regression Classifier,LRC).Both of them have some advantages in mobile terminal.This paper analyzes and studies the algorithms of linear and non-linear regression classifiers,and based on the linear and non-linear regression classifiers,several classifiers are designed.The main contribution of this paper includes the following parts.Based on the idea of linear regression classifier,three improved classifiers are proposed.Those proposed classifiers are focused on the problem of classification imprecision in the classification process.Those proposed methods are distance based a dotted line classifier,false linear regression classifier and distance constrained linear regression classifier.The improvement points of the three classifiers are using the distance information between samples and regression line and spatial characteristics between the sample points,and a large number of contrast experiments show that the new methods can obviously improve the recognition rate of the classifier in face image which is easy to be misclassified under illumination,angle and expression changes.At the same time,based on the kernel function and the nearest feature plane classifier,this paper proposes a kernel nearest farthest classifier and the center-based restricted nearest feature plane with angle classifier.By combining the kernel function with the nearest farthest classifier,the kernel nearest farthest classifier can effectively improve the classification accuracy of the nonlinear separable samples in the original sample space.Besides,by adding angles to restrict the infinite extension of the feature plane,the center-based restricted nearest feature plane with angle classifier solves the problem of misclassification caused by the infinite extension of the feature plane.By comparing the experimental results with other existing linear regression classifiers,the two proposed methods can improve classification accuracy of the nonlinearly separable samples in the original sample space effectively.