Research On Face Recognition Algorithm Based On 2DPCA Of Combined Kernel Function

Face Recognition adopts non-invasive acquisition to meet human beings' idiomatic cognition process that people distinguish others by appearance.With the simple capture device,the technology can fast and conveniently recognize faces in real time.In this case,Face Recognition attracts people's attention and favor among numerous biometric identification technology.In this thesis,a two-way fusion method based on two-dimensional principal component analysis(CK-2DPCA)and two-dimensional linear discriminate analysis(2DLDA)is proposed.Firstly,the two-dimensional matrix samples are mapped into high-dimensional space,and the horizontal feature extraction is carried out by CK-2DPCA method.At the same time,the matrix is extracted by 2DLDA method.This method not only preserves sample's two-dimensional spatial distribution structure,but also solves problem of poor recognition of face image caused by light,expression,gesture and other nonlinear structural factors.The key technical work to implement this improved approach is as follows:(1)This thesis is based on the two-dimensional matrix bidirectional feature extraction analysis method which handles sample images in two directions.And through the introduction of kernel function,it has better performance on dealing with high-order information.The method adopts combined kernel function and maps the sample data into high-dimensional kernel space through kernel mapping.(2)The choice of the combination of kernel functions:because different types of kernel functions has different characters and scope of application,the sensitivity to the data is different.Local kernel function has certain advantages on manipulating data of local information.Global kernel function is better at extracting the global information of the sample,and its generalization ability is stronger.The selection of kernel function leads to different results.In this thesis,we adopt the concept of combined kernel function to select different types of kernel functions according to the demand,and make a reasonable linear combination.This can not only preserve entire sample data,but also highlights the local characteristic among different types to achieve the best results.(3)The choice of the kernel parameter:the selection strategy of parameters has a greater impact on the subsequent recognition effect.In this thesis,we preferentially use the K2DPCA-2DLDA method to select the optimal kernel parameters under the single kernel function.The optimal parameter obtained under single kernel function is then used as a reference in combined kernel function.Then we choose different weight coefficient to perform comparative analysis to achieve the best recognition effect.Through large number of experiments,we conduct research and simulation analysis on the performance of combined kernel function,which proves the superiority of the improved algorithm performance.The face recognition system is designed and implemented by the MATLAB software platform,which proves the practicability of the improved fusion method proposed in this thesis.