Face Recognition Based On SVD-NMF Algorithm

Because of its low acquisition cost,high security and rich application scenarios,face recognition has always been a research hotspot in the field of computer pattern recognition.With the continuous maturity of face recognition technology theory,face recognition as a new type of identity verification technology is widely used in various practical scenarios.However,due to the complexity of the actual scene environment,the actual face recognition effect is also limited.Among them,in the face of massive face image data,how to train effective feature information is the key to face recognition technology.Non-negative matrix factorization(NMF),a classical algorithm for feature extraction,has been widely used in face recognition research in recent years.How to speed up the convergence speed of matrix decomposition and make the decomposition result more sparse is an important direction to improve the NMF algorithm.Combining the singular value decomposition(SVD)algorithm with the NMF algorithm is applied to face recognition,and the effect is better than a single algorithm.At present,there are many corresponding SVD-NMF fusion algorithms,such as the CSVD-NMF fusion algorithm and the SVD-NMF algorithm.This paper mainly focuses on the improvement of SVD and NMF fusion algorithms and their research and application in face recognition.The main contents of this article are as follows:(1)Based on the CSVD-NMF algorithm,this paper proposes two improvements to the shortcomings of the CSVD-NMF algorithm.1)This paper proposes a new method to calculate the contribution rate of singular values,and then determine the optimal number of singular values in SVD.Compared with the previous calculation method,the method is more reasonable in intercepting the optimal singular value in the same reconstruction error.2)This paper proposes an improved method of initializing NMF.When the rank of the original matrix is large,the relative distance error between the original image and the reconstructed image can be effectively reduced compared to the NNDSVD and SVD-NMF algorithms.Based on the above two improvements,the improved CSVD-NMF algorithm has better sparsity and relative distance error after decomposition,and the effect of applying on face recognition is better.(2)This paper proposes a fusion algorithm based on CSVD-BFBNMSF.1)Because the feature description ability extracted by NMF algorithm is not high enough when processing matrix dataset,the generalization is very poor,and BFBNMSF algorithm can effectively solve this problem.Therefore,based on the CSVD-BFBNMSF algorithm proposed in this paper,the extracted feature description ability is better than the CSVD-NMF fusion algorithm.At the same time,the actual effect of the algorithm is better when dealing with small sample learning problems.2)This paper initializes BFBNMSF with SVD,which has more stable experimental effect and faster average convergence speed than random initialization of BFBNMSF.Based on the above two points,the proposed algorithm based on CSVD-BFBNMSF is better in face recognition.