Algorithms Research On Face Recognition Based On Locality Preserving Projections

Feature extraction is one of the most basic problems in the field of facerecognition. At present, the importance of face feature extraction is focused onthe feasibility and effectiveness of the algorithm. Although there are manyclassical algorithms to solve it, these algorithms have been made a great success;but the research shows that face samples are located in low-dimensionalnonlinear manifold which is embedded into high-dimensional image space,however, the traditional subspace linear analysis methods cannot describe themanifold structure accurately and efficient.Locality preserving projection (LPP), as a linear feature mapping, can betterreflect the manifold structure of the samples, but it still is an unsupervisedlearning method, with weakness of the classification, not the most effectivealgorithm to the face recognition. In view of this, the paper gives three kinds ofsupervised feature extraction algorithms, which are based on the LPP algorithm:1. On the basis of the CCA algorithm and LPP algorithm, the paperproposed the RLPCCA algorithm. The algorithm will effectively combine CCAalgorithm LPP algorithm by introducing the type of information, not only be ableto maintain the partial information structure in the sample, but also to maximizethe information between the two sets of samples.2. On the basis of the MSDC algorithm, the paper proposed the UMLPPalgorithm. The algorithm will search an optimal set of unrelated set ofdiscriminant vectors, not only to eliminate the redundancy between features tofacilitate the reconstruction of the data, but also makes the projector maximumand minimum degree in class to improve the algorithm’s validity and stability inthe feature space between the sample classes divergence.3. Multiple locality preserving projections (MLPP) was based on locality preserving projections. When constructing the graph, each point directly relatedto the same as k neighbor point, identified close neighbors point at the same time,also identified a close neighbor points. It not only retains the stability of the localstructure, but also maximizes the sample class, with full consideration ofdifferences between manifolds. The similar points with each other closer, whichcan form a high clustering, which made the thedifferent spots away from eachother, and the same point to approach each otherThe feasibility and effectiveness of algorithms had been demonstratedthrough experiments conducted on face databases.