Research On Face Recognition Methods Based On Sparse Representation And Subspace

The key issue of a successful face recognition approach is how to extract discriminant features from a face image and recognize the target by the classification criterion. This dissertation focuses on the feature extraction technique and classifiers and pays more attention to the most popular technology for feature extraction-subspace analysis and the research hotspots-sparse representation in face recognition. Several efficient feature extraction and recognition algorithms have been proposed. The main contribution and innovative points of the dissertation are summarized as follows:1) In the locality preserving projections (LPP) algorithm, the false close neighbor relations is still maintained after projecting data to the subspace, moreover, reconstructing the overall inherent laws by the approach of partial mergers can not better reflect the distribution of differences. To address these problems, maximizing margin and discriminant locality preserving projections(MMDLPP) is proposed, which can find and reconstruct the local intrinsic geometric structure and the overall distribution difference of data sets. To keep the real close neighbor relationship, separate false close neighbors from sample data and expand the distance between sample and its false close neighbors and the distance between different types of samples, the method has great discriminating power.2) Sparsity preserving projections (SPP) algorithm does not well reflect the discriminant information when calculating the sparse weight. To handle these issues, discriminant sparsity preserving embedding (DSPE) is proposed, which can reconstruct the local intrinsic geometric structure while maintain the characteristics of sparse representation and obtain the low-dimensional global optimal embedding. DSPE is a linear supervised learning method which can extract features effectively and has good robustness. 3) Nearest subspace classifier (NS) and sparse representation based classification (SRC) provide a good recognition performance in face recognition. However, as reconstruction-based classifiers, they inevitably need a lot of reconstruction operations of a sample. To overcome the disadvantages of expensive computation cost, an improved version of NS and SRC are proposed. A novel space called reconstruction space is constructed by the reconstruction proportions. The point in the reconstruction space denotes the case of a sample reconstructed by training samples. The improved NS and SRC seek to find an optimal mapping from the conventional sample space to reconstruction space by the prior knowledge. Then a new sample after mapping to the new reconstruction space would be classified quickly by the reconstruction proportion in reconstruction space. The improved classifiers can outperform NS and SRC in recognition speed and classification accuracy. What’s more, this framework can be better extended to other reconstruction-based classifiers.