Research Of Uncorrelated Fetaure ExtractioN Algorithm

Feature extraction is the elementary problem in the area of pattern recognition .It is attractingmore and more people to research .With the development of the technology ,the methods of featureextraction are coming out in an unending flow. Uncorrelated optimal discrimination vectors is oneof them.The conventional method to solve the eigenvalue is iteration .But with the increasing of thetraining samples number, computational complexity increase quickly ,which is the maindisadvantage of the algorithm .As we know, although the single discriminant vector can make theFisher criterion maximum, the vector set we get in the end maybe not the optimal. So in this paper,based on the generalized singular value decomposition, we propose a new algorithm ,namedGSVD-UODV .We can not only avoiding the small sample problem but also get the vector setdirectly through this method. Furthermore, our method can reduce the computational complexity. Inthe end , after we got the vector set ,we will use the totality scatter matrix which can make thevector set uncorrelated .In order to make the algorithm used in other methods , we have extended the algorithm in thenonlinear space . This section is in the fourth chapter , named GSVD-KUODV . We also give thetheory and the formula detailed .In the last section , we do our experiment on the AR database and the CAS database .Wecompare the performance and the computational complexity .The results show that our approachgives the better performance compared with the UODV with much smaller computationalcomplexity.