Research On Dimensionality Reduction Of Face Images Under Uncontrolled Environment

Face recognition is one of the most widely used techniques for biometrics,and it is one of the most important research topics in the fields of information security and machine vision because of its convenience,intuition,practicality and friendliness.With the rapid development of the Internet of Things and sensing technologies,face images collected in real environment exhibit the highdimensional,non-constrained and multi-modal characteristics that lead to complex and changeable data and the nonlinear distribution in high-dimensional space.In order to solve this problem,thesis studies the dimensionality reduction of high-dimensional and complex non-constrained face images,and expects to extract the low-dimensional essential structure to characterize different face and improve the accuracy of face recognition.The specific work contents are as follows:(1)Sparse Preserving Projection(SPP)is an unsupervised dimensionality reduction method that uses sparse reconstruction of all samples except for the to-be-expressed sample,without considering the sample's category information.To solve this problem,thesis proposes a new method,named Discriminant Sparse Preserving Projection(DSPP)algorithm,which only uses the remaining samples of the same category to reconstructed.Based on the compact constraints of the samples in the class,the least square method is used to realize the quick solution of the reconstruction weight values.Besides,in the low-dimensional projection,not only the reconstruction relation of the samples,but also the global distribution are taken into account in order to achieve the best projection of the highdimensional data.The experimental results show that this algorithm can fully guarantee the reconstruction of the same category samples,avoid the interference of heterogeneous samples and effectively improve the accuracy of unconstrained face recognition.(2)In view of the DSPP algorithm only using the same category samples to reconstruct the represented sample,without considering the reconstruction effect of heterogeneous samples,thesis proposes Discriminant Sparse Preserving Embedding(DSPE)algorithm on the basis of DSPP.Firstly,the intra-class reconstruction weights and inter-class reconstruction weights of the samples are respectively calculated,and then the intra-class dispersion and inter-class dispersion of the samples in the low-dimensional space are respectively calculated by using the reconstruction relations of the samples and the global discrimination information.Finally,the best projection dimensionality reduction of high dimensional data is achieved by minimizing inter-class dispersion and maximizing intra-class dispersion of samples simultaneously.The experimental results show that,compared with the DSPP algorithm,the DSPE algorithm not only considers the similarities of the same category samples,but also takes into account the differences of heterogeneous samples,and combines the local reconstruction relations with the global discriminant information to obtain a better discriminant lowdimensional subspace.Unconstrained face recognition accuracy has been significantly improved.(3)Considering that the DSPP algorithm is a dimensionality reduction method based on vector representation,it not only destroys the spatial structure information of the image,but also generates the dimensionality disaster and the small sample problem.Considering this defect,thesis also extends the DSPP algorithm to tensor space,and proposes Discriminant Tensor Sparsity Preserving(DTSP)algorithm.The algorithm directly reconstructs the tensor data,which is more conducive to accurately describe the neighborhood relations and spatial structure of the samples in high-dimensional space,and makes the low-dimensional subspace more discriminative through bilateral projection.Experimental results show that the DTSP algorithm has a higher recognition rate than the dimensionality reduction method based on vector representation.