Research On Feature Representation And Classification Method Based On Data Internal Geometry

According to the hypothesis of manifold,high dimensional data has the geometry structure of low dimensional nonlinear manifold,that is manifold.On this basis,in order to describe the low-dimensional manifold structure in high-dimensional data better,this paper describes the inherent principle of robust data modeling by using the method of contraction CAE.It can describe a low-dimensional manifold structure in the input data space.We find the low-dimensional manifold depends on the sparsity of singular value of Jacobian matrix and propose a new feature representation method-strong contraction auto-encoder(TCAE).By using the constraint of trace norm,the Frobenius norm constraint of in CAE is changed to trace norm constraint,which makes the singular value corresponding to Jacobian matrix more sparse.It is helpful to learn a manifold with lower dimension and denser,so as to describe the local low dimension properties of the manifold more robustly.It is proved that the optimization problem corresponding to TCAE has a closed form solution.Based on the primal-dual algorithm,the original optimization problem is divided into two subproblems to solve TCAE.And we proved the convergence of the algorithm.In order to verify the application effect of TCAE,this paper takes Momoge wetland as the research object.Select Momoge wetland's seven bands and56 texture features as the features of remote sensing images.It can be divided into six types: residential area?water area?farming area?grass area?wetland area and unused area.Extract features by TCAE and CAE respectively and classify the features by the SVM.We obtain the classification performance indexes and kappa coefficients of TCAE and CAE.The experimental results show that the features extracted by TCAE are better than CAE in classification effect.It proves that the features represented by TCAE are more robust.