Face Recognition Algorithm Research And Parallel Implementation Based On Tensor Subspace

Face recognition is a research hotpot in recent years, which is in the field of pattern recognition, and its scope of application is also expanding. The traditional algorithms used for face recognition need to translate the2D image into a vector, then analyze the vector as original feature. But there are many faults and shortcomings:Firstly, the space relationship between the image pixels after vectorization is most likely damaged; Secondly, the dimension of the statistical parameters leads to high computational complexity and storage costs; Finally, the number of samples is much smaller than the dimension of the vector usually, which has the problem of singular covariance matrix.Face recognition algorithm based on tensor space is to be considered from the image feature, which do not need to translate the image into a vector and deal with2-order tensor. It can keep the spatial correlation, reducing the dimension of the covariance matrix. Meanwhile, it avoids the problem of small sample size and improves the performance of face recognition.This paper studies and implements Tensor Principle Component Analysis and Tensor Linear Discriminant Analysis based on tensor space. Furthermore, as to the feature extraction is not fully, the recognition performance is poor in the low dimension of TensorLDA, we proposed an iterative approach to overcome the phenomenon. Using ORL and Yale human dataset to test the performance of algorithm. The experiments show:In ORL dataset the average recognition rate of TensorPCA is1.713%higher than PCA; It-TensorLDA is1.88%higher than TensorLDA and3.03%compared to Fisherfaces. In Yale dataset the average recognition rate of TensorPCA is1.3%higher than PCA; It-TensorLDA is0.91%、3.14%higher than TensorLDA and Fisherfaces.There is a problem that Matrix multiplication and using Jacobi to obtain eigenvalues and eigenvectors consumes a lot time during the implementation of algorithm. In order to improve the speed of algorithm, the thesis adopts OpenMP and SSE to parallel the algorithm. On the platform of Intel CoreTM2computer with two cores, the result shows:The speed of TensorPCA using multi threads is1.407times and SSE instructions is1.583times faster than serialized version. The speed of TensorPCA using OpenMP and SSE is1.869times. Compared to TensorLDA, The parallel TensorLDA using multi threads is1.572and SSE instruction is1.167times faster than serialized version. TensorLDA using OpenMP and SSE is1.856times faster than serialized version.