Research About Algorithm Of Multi-view Subspace Clustering Based On Block Diagonal Representation

With the rapid emergence of data that can be described by different feature sets or different“views”,multi-view subspace clustering has attracted considerable attentions in computer vision community.Recently,many multi-view subspace clustering approaches have been proposed.To handle the disadvantages of them,this article takes advantage of block diagonal representation and cauchy loss function.And two multi-view subspace clustering methods are developed,which are summarized as follows:?1?Many existing methods exploit multiple views information by virtue of the advantages of sparse or low-rank representation.We can observe that these methods own the common block diagonal property which potentially leads to correct clustering.While these methods are indirect,they could capture the block diagonal property under the independent subspace assumptions.However,in real applications,due to the data noise or outlier,the required assumptions often do not hold and the block diagonal structure is violated.Multi-view subspace clustering via block diagonal representation is proposed to deal with the above problem.In contrast,the usage of block diagonal representation not only directly encourages the matrix to be block diagonal,but is also able to control the number of blocks,which is important for multi-view subspace clustering.For each view,the subspace representation matrix is learned in a self-expressive way.And we directly learn a consensus matrix which represents the underlying subspace structure of the multi-view data,by using the block diagonal representation.What's more,the experimental results show that our method could improve the accuracy of clustering.?2?To reveal the common latent structure shared by multiple views,existing methods usually impose the sparse or low-rank constraint and use Frobenius norm or?₁-norm to measure the reconstruction errors of multi-view data.However,the assumptions of sparse or low-rank representation works are difficult to satisfy in real word.Besides,the Frobenius norm or?₁-norm is suitable to handle either Gaussian noise or sparse noise,which are very sensitive to large noise or outliers.When the data is contaminated by large noise,performance of existing methods are degraded dramatically.In this article,we propose a novel multi-view subspace clustering method based on block diagonal representation and cauchy loss function.The method learns the consensus matrix by assigning suitable and different weights to different views.The setting could fully take advantage of multi-view information and alleviate the influence of some unreliable views.And our approach is more straightforward and robust as the application of block diagonal representation and cauchy loss function,respectively.Finally,the experimental results demonstrate our method's effectiveness.