Research On One-class Target Classification Based On Covering Model Of Space Structure Of Manifold

One-class classification problem is one of the hot spots of the modern field of patternrecognition, and be widely used in practical applications such as identification andanomaly detection. These problems can’t use the distinction information between differentmodes of class, but the commonality information among the same mode. It needs to buildan effective covering model to complete classification task. This paper do some researchon one-class classification based on the achievements of previous researchers, and proposesome novel one-class classification covering models.Firstly, the cover radius of Minimum Spanning Tree Class Descriptor is fixed, whichmakes local resolution the same. In this case, this paper proposes multi-analysis scale theminimum spanning tree covering model. Consider close neighbors distribution and localedge length ratio on the basis of the minimum spanning tree model to complete the pointscale and edge scale analysis. The two scales integrate into the scaling factor as base forcoverage radius to achieve the dynamic change of the different regions of the resolution.Experimental results show good performance.Then, traditional modeling algorithms can not describe the manifold internalstructures and the local edge in details in the same time. This paper proposes closed-looptree with minimum mean value covering model.It uses minimum balanced tree to coverthe main structure of data and packs the branches in edge area with closed loops.Thiscovering model can genarate a smooth and specific edge by the work. This method holdsconnectivity and omnipresence of manifold and uses different models and distinguishscale to describe the backbone structure and edge detail. The experiments based on severaldatabases prove that the proposed manifold is effect in data covering description.Lastly, there are lots of noise and redundancy attribute in data original space, itsintrinsic manifold prone to be distorted or even deformed. This paper proposes one-classclassification based on multi-scale 1-norm constraint subspace. Model and classify inlow-dimensional subspace of the sample. The combination of multi-scale 1-norm Graphand asymmetric LPP algorithm preserves the original local neighboring structural features in the subspace. And build neighboring hypersphere model based on the 1-norm distance.Experiments on several databases show good performance.