Research On Face Recognition Based On Hybrid Kernel Function SVM And Genetic Parameter Optimization

The technology of face recognition has the advantages of non-contact,convenience and reliability,and is widely used in the financial security,traffic,information security and other fields.Although there have been many breakthroughs in the research of face recognition in the past few decades,many problems still need to be solved.For example,facial expressions,posture,illumination and other changes would cause the decline of recognition rate.As a consequence,the research on the technology of face recognition has a profound historical significance.As a classical Machine Learning method,the kernel function based on Support Vector Machine and the choice of parameters in kernel function has a significant influence on the performance of classifier.The use of kernel function technology not only greatly reduces the amount of computation in the input space,but also improves the classification performance of Machine Learning effectively.In the field of Machine Learning,the choice of kernel function and the construction of kernel function are a difficult task.However,there are not many contributions in this field.In the view of problems mentioned above,this paper discusses the construction of SVM kernel function in detail,and focus on the improvement of the traditional parameter optimization algorithm.This paper starts with the structure of the system of face recognition.Firstly,the basic image preprocessing technology and principal component analysis are given.Then,this paper introduces the SVM and the knowledge of statistical learning theory,as well as the concept of VC dimension,generalization and empirical error.The kernel function method has been deeply discussed,including the concept of kernel function and Mercer condition theory.Additionally,this paper also introduces the common expressions of polynomial kernel function for example the Gaussian kernel function,the linear kernel function and the Sigmoid function.The disadvantage of Gaussian kernel function is the poor global promotion ability but its learning ability is extremely strong.In the contrary,the polynomial kernel function,as a kind of global kernel,has the better generalization ability than Gaussian kernel function,but its learning ability is not as good as Gaussian kernel function.Considering the characteristics of Gaussian kernel function and polynomial kernel function,this paper combines the two kernel functions a mixed kernel function.The mixed kernel function satisfies the definition of the kernel function and the related theorem,and proof is given.Then this paper introduces the traditional parameter optimization algorithms,including grid search,genetic algorithms and cross-validation algorithm.However there are some defects in these optimization algorithms.Genetic algorithm is characterized by fast searching speed,but the optimal solution is not accurate enough.For the grid search method,it takes much time to search the optimal solution,but it gives the accurate optimal solution.In response to the above problems,firstly this paper uses the heuristic algorithm of genetic algorithm to find the approximate range of the optimal solution quickly,and then uses the grid search method in the range of the second accurate search.The improved algorithm can not only shorten the computational time of grid search method,but also has a better solution than the genetic algorithm.Through the simulation experiment,the improved hybrid kernel SVM and parameter optimization algorithm are applied to the ORL face database,which greatly improves the recognition rate,and experiments on the image in the face database after adding noise to verify the practicability and robustness of the algorithm.