The Research Of Clustering Algorithm Based On Independent Subspace Analysis Networks

Independent Subspace Analysis(ISA)has a very effective nonlinear feature extraction ability and has been widely applied in face recognition,image segmentation,image understanding and image clustering.In these practical applications,the key of them is to extract the valid features of the data.Many of the existing methods of extracting features are relatively straightforward and seldom use the original structure information of the data.Based on independent sub space analysis network,this paper proposes a deep network clustering algorithm combining sparsity prior information of data.This algorithm take the data prior subspace information obtained by training as the output constraint of independent subspace analysis network and is used for cluster analysis.In this paper,two methods are proposed for single layer and multi-layer subspace analysis network learning.This paper introduces the structure information of original data into the model,and proposes a network called Independent Subspace Analysis with Sparsity Prior(ISASP).The ISASP network obtains the feature vectors of the data by mapping the data nonlinearly from the input space to the feature space,and makes the feature vectors have a similar subspace structure with the original data.The feature vectors are used to construct the affinity matrix of the data,and then use the spectral clustering algorithm to obtain the clustering results.ISASP takes the sparsity prior information of data to constrain the network output,so that the model can be more inclined to give the optimal feature vectors of the data.In this paper,the results of clustering experiments on two benchmark data sets are compared with other algorithms to verify the performance of ISASP.In order to solve the problem that the efficiency of ISASP model will gradually decrease with the increase of the feature dimension of input data,we propose a network called Stacked Independent Subspace Analysis with Sparsity Prior(Stacked-ISASP).The algorithm divides the data into several parts and input them to multiple ISASP subunits to reduce the parameter complexity of the network.This paper validates the performance of the algorithm by using the experimental results on the four public datasets,CMU-PIE,ORL,COIL20 and USPS.In the aspect of clustering,this paper introduces several common clustering algorithms.By the study of these algorithms,the sparse subspace clustering algorithm based on spectral clustering is selected as the final division of data.By using the essential feature learned by Stacked-ISASP model from the data,we completed the experiment of five evaluation indexes on four datasets.The algorithm proposed in this paper is proved to be effective and has achieved good results on all indexes.Specifically,the experimental results were as high as 88.42%,86.35%,98.51%and 98.69%respectively on average clustering accuracy.