Intelligent Dimension Reduction Technology Research And Application

The technology of pattern recognition is playing a more and more important role in the following fields:character identification, speech recognition ,biomedical and other areas. With the development of computer technology, the problems that computers faces are becoming more complex. We often come across various high-dimensional datas in life,such us biological gene sequencing data,stock data,satellite remote sensing data,internet multimedia data and so on. Maybe there are two problems when computers process the datas.First, the"dimension disaster", which makes it hard to calculate and deal with,or impossible to handle. Then it will take unprecedented challenges to the pattern recognition and the rule exploration of high-dimensional datas; Second, there will be a lot of redundant information remained if the dimension is too large,and it not only affects the performance of the computer processing but also brings interference to pattern recognition . How to find the appropriate low dimensional representation of the high-dimensional data based on certain rules, the resulting internal structure information and good methods to deal with it become one of the hotspots researchers work on. Dimension reduction methods draw great attention as a means of solving high-dimension disaster and the corresponding exploration is endless.In this paper, after studying many domestic and foreign algorithms related to dimension reduction, we found there are some defects in the existing algorithms, such as the algorithm itself is not flexible enough or the effect of application areas is not good enough. The paper proposed improved algorithms and verified the effectiveness by lots of experiments.The main work of this paper was summarized as follows:The first section is introduction. In this section we give a brief description about the current research and application areas of the pattern recognition technology and a special study on face recognition. Then we make a review of the existing theory and algorithms of the smart dimension reduction technology.In the second section, we introduce the latest research results of dimensionality reduction algorithms, then we focus on the Local Preserving Projection and Approximately Harmonic Projection proposed by HE Xiaofei and his partners .In the end, we describes the application of dimension reduction in face recognition and clustering areas.In the third section, we introduce index parameter to the diagonal matrix of the Local Preserving Projection algorithm, we call it An Improved Local Preserving Projection Algorithm Based On Exponential Diagonal Matrix.Through experiments it can be proved that the algorithm can affect the results of the dimension reduction and make it easier to approach the intrinsic dimension of the datas.We also check the discrimination after reducing the dimensions and the sensitivity of noise.In the fourth section, improving on the constraint of the Local Preserving Projection algorithm, is made and we call it Local Preserving Projection Algorithm With the Exponential Constraint. This algorithm adjusts the constraint used in the solving of the objective function.It enhances the flexibility in solving problems. After adding the index we make a lot of experiments to observe the influence of dimension reduction and the rate of recognition. In addition, we attempt to summarize the scope of the parameter and design experience.In the fifth section,we make improvement on the Approximately Harmonic Projection algorithm by adding two index parameters into the constraint , which can generalize the formula we want to solve .Through face images clustering experiments, we find that the change of the index can make a great impact on the result of the clustering.We can adjust the parameters to achieve better clustering results aiming at certain face clustering.