Reserch On Face Recognition And Image Retrieval Methods Using Subspace Analysis

In the content-based image retrieval system and a face recognition system, feature extraction play a basic role to complete the retrieval and recognition task as an important link. This article investigates the subspace analysis method for a deeper research, and on this basis offset the shortage of the algorithm, and some more effective feature extraction methods are proposed. And ,these methods are verified to be effective in the application of retrieval and face recognition system.For subspace analysis technology with the retrieval and recognition application, launched the following job:1. Using the subspace analysis to apply to the shape based image retrieval.As shape feature descriptors,high dimention zernike moments have the function of describing the detail information of image region,which exist "dimension disaster".This will result to increase the complexity of the algorithm and unnecessary information which make major information confused,and will affect decribing the content of the image. A new algorithm which based Locality Preserving Projections(LPP) Manifold method proposed to realize dimension deduction in image data. Under the condition of Laplace figure keeping local sample data,while overall algorithm is introduced to ensure the integrity of the sample. Considering the influence of the correlation between information to affect projection accuracy. To get to the eigenvector with Schur eigenvalue decomposition to obtain the orthogonal vectors.This can make the data reconstruction relatively easier, and the rotation invariant of Zernike moment can still keep down, then making the image retrieval accords with the human visual effect. This method is superior then LPP in the retrieval performance.,and retrieval results have significantly improved.2. It is mainly based on the original gray-scale face for feature extraction,using a statistic learing algorithm-linear discriminant analysis(LDA) to have a study of the subspace. After the careful study with the traditional definition of a matrix,found that when the two adjacent classes have the similarity,it is easy to generate an error and result in unclear data classification. Proposed a new subspace learning method---largest margin neighbor dicriminant analysis(LMNLDA).It accords the neighbor criterion to project the datas to a seprarable subspace,and refreshes the original scatter matrix in this new space to construct a new target function.To overcome the drawback that the traditional definition of matrix for the mean value of two kinds scatter matrix or more categories that unsimilar to distnguish between samples.And on this basis it solves the small sample problem .Taking a expriment on the standard face database to prove that effectiveness of the results.