The Study And Application Of Feature Extraction Technology With Local Learning Ability

In recent years, the importance of the pattern recognition technology in the science and technology innovation is growing rapidly. The feature extraction technology as a key link in the process of pattern recognition has also got more and more researchers’attention. With the development of science and technology, the large amount of information we have been collectted cause the information explosion. Therefore, feature extraction technology as a effective way to sovel "dimension disaster" has been studied widely. However, some traditional feature extraction methods, due to a series of problems, such as small sample size problem and nonlinear problem, etc, can’t get very good result of the feature extraction. Based on the analysis, some feature extraction technology with local learning ability has been put forward in this papge. the main research contributions have been given as follows:LA linear feature extraction method with local learning ability has been proposed:local sub-domains based maximum margin criterion, LBMMC. This method is based on the LDA, indicate when the small sample size problem occurs, LDA will fail to work. To solve the small sample size problem effectively, we bring the maximum margin criterion in this method, change the ratio of the scatter to difference value. What’s more, we use the local weighted mean to instead of the mean standard of LDA to remain the local information.2.The nonlinear feature extraction methods have more advantages on exposing the nonlinear structure embedding in the sample. Based on the KFDA, anew nonlinear feature extraction method is proposed:locally weighted based nonlinear feature extraction method, LWNFE. This method can reflect the local nonlinear structure of samples well.3.With the influence of the matrix pattern, a linear feature extraction method with local learning ability has been proposed:local sub-domains maximum margin criterion based matrix pattern, Mat-LSMMC. This method is based on the MatFLDA and combine with the concept of the local weighted mean, turn the vector form of sample into matrix, what remains the information between ranks.4.Based on the concept of the nearest feature line, the nonlinear feature extraction method based the nearest feature line embedding has been proposed, NFE-NFL. Due to the conbination with the nuclear technology and the laplace graph between the feature lines, this method has the better classification results.