Dimension Reduction Of High Dimensional Data With Application To Image Recognition

New technical revolution, especially the flourishing development of information technology, has brought unprecedented opportunities for the prosperity of the society.This is an era of whose characteristic can be digital where data become valuable wealth and resources in various fields. Industries produce a large number of high dimensional data every day,and these data present informative,diversify,complex structure and so on. Rich data not only include a lot of meaningful data, but include worthless ones as well. Enormous dimensions conceal their intrinsic, and bring interferences and great difficulties to data analysis and processing. Using dimension reduction technology to analysis data in a scientific way for obtaining information that users are interested in is an important part of scientific research.Dimension reduction reduces the dimensions of high dimensional data. What's more,it helps to obtain a more concise and effective representation with low dimension. It conducts transform from high dimensional space to low dimensional space by learning a mapping mechanism and maintains meaningful structure. In this paper,deep researches are conducted to the method, theory of dimension reduction technology and its application in practice, and it propose new methods base on these researches and obtained some success. The main work of this thesis includes the following:1. This paper briefly reviews dimensionality reduction technology about the background, purpose and significances. Also, it summarizes the status and development in domestic and international, and it demonstrates the characteristics of linear, kernel and manifold learning methods by existing examples.2. This paper introduces dimension reduction algorithms and the work of preprocessing of image containing series operations prior to the implementation of image dimension reduction. Preprocessing is vital as to get desired form of images preparing for the dimensionality reduction and recognition.3. A sparse representation classifier based discriminant projection method is proposed, which match the classifier in building discriminative model to improve the precision. Sparse representation illustrates the similarity among samples. So,according to it, the proposed method generate dimensionality reduction model to optimize two targets. One is sparse reconstruction residual of between-class and within-class of the data set, the other is the data distinction. The optimization leads to a higher accuracy of sparse representation classifier when the samples are projected into low dimension space. Face recognition experiments were conducted on AR and Yale databases, and the results show the effectiveness and robustness of the proposed method.4. Considering that global methods lack of useful discriminant information and local ones exist defects in measuring neighborhood relationships,a novel dimensionality reduction method, named margin discriminant projection, is proposed.This method defines new criterion for margin, and it optimizes the margin so as to maximize the within-class similarity and minimize the similarity of between-class.Recognition on two expression databases shows a more satisfying result of the proposed method than the traditional ones. The proposed method can extract more distinguishing low dimensional features, and effectively improve the accuracy of facial expression recognition.