Structure Preserving Local Coordinate Coding

In the field of pattern recognition,local representation is a new feature representation.Manifold learning is an effective dimension reduction method.Due to their effectiveness many researchers give attention to them.On the basis of local coding and manifold learning theory,we proposed structure preserving local coordinate coding model.The main innovations of this paper are as follows:1)This dissertation proposed a dictionary which maintains manifold structure.Traditional method used an iterative approach to learn dictionary or put original data as a dictionary directly.Using these methods to learn the dictionary does not reflect the essence of the original data.To solve this problem,we use manifold learning method to learn dictionary that embodies the essence of the original data.2)This dissertation proposed a local representation.The signal source is represented by all signals directly that does not distinguish the sources of signals.As a result the base signal has not been utilized fully.This paper proposed a method of local restrictions that aims at using less atoms to represent the original signal better.It is coded by nearby date points that remains the neighborhood relations.This dissertation proposed a mathematical model that contains local coding and manifold Learning.Compared with some present dimensionality reduction methods,the accuracy of the structure preserving local coordinate coding is higher.So our algorithm has more value.