Abstract: Locality preserving projections (LPP) aims to preserve local structure of the data by constructing a nearest-neighbor graph.In the construction process of nearest-neighbor graph,LPP will encounter the difficulty of the selection of two parameters K and σ.The construction of nearest-neighbor graph plays an important role in recognition effect,so selection of the two parameters can affect the discrimination ability of LPP.In order to avoid the effects of the selection of parameters on recognition rate,an face recognition algorithm based on improved adaptive locality preserving projection was proposed.Firstly,a parameter-free graph construction strategy is designed,which can adaptively choose neighbors of each sample point and determine corresponding edge weights.Then,because of the high dimensionality problem in the matrix calculation process,QR decomposition is used to reduce dimension.Finally,conjugate orthogonalization is used to reduce the statistical correlation between feature vectors and improve the recognition rate by ensuring that the projection axis has statistical uncorrelation.The experimental results on ORL database show that the new algorithm is better than the LPP algorithm,DLPP algorithm,and LMMC algorithm in terms of recognition rate.

Key words: Nearest-neighbor graph,Adaptive locality preserving projection,Face recognition,Conjugate orthogonalization,Statistical uncorrelation,Dimension reduction