Dimensionality Reduction Method Based On Manifold Structure And Its Application In Face Recognition

In the wave of machine vision and pattern recognition research,face recognition technology has broad application prospects in daily life such as social public security,identity verification,and monitoring,due to its universality,security,uniqueness,and stability.Dimensionality reduction is the most effective method to solve the "dimensional disaster" in face recognition and is highly concerned by people from all walks of life.Because the face image is vulnerable to many external factors,the current dimension reduction methods do not meet the standards of many applications.How to better maintain the manifold structure after high dimensional data mapping is the primary purpose of dimensionality reduction.This paper summarizes and analyzes the existing dimensionality reduction methods.Based on this,three new dimensionality reduction methods are proposed and a large number of experiments are used to verify the effectiveness of the new method.The specific study is as follows:(1)The accuracy of the local dimension reduction method is directly determined by the structure of the graph.This paper focuses on the problems in graph construction,such as the selection of neighboring parameters,the sensitivity of noise,the lack of discriminatory power,and the inability to merge fields.Through the in-depth study and study of graphs,the traditional K-nearest neighbor graphs are improved and the boundaries are redefined.A local dimension reduction method based on sparse average boundary graphs is proposed.The sample average boundary is used as a similarity measure to construct neighbor graphs,and the K neighbor selection is reduced.Loss of performance due to K value and simplifying neighbor graphs by sparse representation.Compared with the previous method,this method can reduce the computational complexity,shorten the calculation time,and increase the recognition rate.(2)Compare the common problems of several distance measures.A global dimension reduction method based on arc distance is proposed.The new arc distance distance is used as a new distance metric for dimension reduction method.The radian and radius are used to measure the difference between the values and directions in the vector,and the shortcomings of existing distance metrics are compensated.Experiments show that this method can improve the performance of data classification,improve the recognition rate,and provide convenience for subsequent face recognition.(3)Because the linear dimensionality reduction method can intuitively represent the mapping relationship and the computation amount is relatively low,it is widely used in face recognition.However,most real-time facial images are not linear,resulting in a linear dimension reduction method that does not yield very good performance.Based on this problem,a fusion-based dimension reduction method based on kernel is proposed.The nonlinear face image is mapped to high-dimensional space by kernel function to achieve linear separability.Then the global and local dimension reduction methods are combined to reduce dimension,and the data is fully utilized.Manifold structure and category information get better recognition rate and robustness.