Sparse Representation Model And Its Application

Nowadays,we are in an era of big data.High dimensionality of big data,brings great challenge in using,processing,and analyzing big data.Sparse representation is an effectively method for representation of high dimensional data.The basic idea of sparse representation is to represent high-dimensional data with a linear combination of as few bases or dictionary atoms,thus characterizing the low-dimensional nature of the data.Subspace clustering is an aims to segment high-dimensional data into clusters,each cluster corresponds on subspace.It can effectively reveal the potential subspace structure of the data.In thesis,we study the sparse representation of high-dimensional data and its application in subspace clustering,mainly considering the correlation between data and data,the global structure of the data and the intrinsic geometry of the data,a more efficient subspace clustering model was designed.The clustering performance of the model was verified on the commonly used datasets,and the given model was applied to face recognition and image segmentation.Our main contributions include:First,a new sparse subspace clustering model is presented by using a regular term with the correlation of the data.The model is adaptive and helps to force related data to be clustered together and unrelated data to be separated.Second,a model of joint learning data similarity and clustering results is presented.The nuclear norm is introduced to characterize the global structure of the data,and the grouping effect is defined to capture the intrinsic geometry of the data.Model update iteratively the representation coefficient matrix and the segmentation matrix,until the segmentation matrix no longer changes.In the iterative process,the segmentation result obtained after iteration has better precision because the representation coefficient matrix has the effect of uniform inter-class sparseness within the class.Extended experiment are carried out on the data sets commonly used in several cluster analysis.Experimental results show that the proposed model has better clustering performance than other related methods.