Multi-view Tensor Analysis And Deep Clustering Analysis In Unsupervised Learning

Unsupervised learning is an important research topic in computer vision.Image clus-tering,as a typical application in unsupervised learning,has received extensive attention and rapid development in recent years.Among the traditional image clustering methods,the subspace clustering based on multi-view tensor analysis improves the multi-view clustering performance to a higher level.However,the previous multi-view tensor analysis algorithms were developed according to the linear assumptions in the low rank representation theory.Thus,when real data is sampled from multiple nonlinear subspaces,the performance of these algorithms is limited.In order to tackle it,this thesis studies the problem that the subspace clustering method based on multi-view tensor analysis cannot deal with nonlinear conditions in image clustering.In the meanwhile,there is also a multi-view fusion problem in image retrieval which is another commonly used unsupervised learning application.In the past,it is difficult to capture high-order information between indexed views,resulting in insufficient utilization of multi-view indexes.Aiming at this problem,this thesis explores whether multi-view tensor analysis method in image clustering can be used in image re-trieval.In addition,deep learning has shown superior performance in various fields,but there are few studies on deep learning for clustering at present.What's more,it is difficult to embed a small amount of supervised information in the deep clustering framework to further improve clustering performance.This thesis analyzes the framework of deep cluster-ing.In summary,this thesis studies multi-view tensor analysis and deep cluster analysis in unsupervised learning.The research content and main contributions include the following aspects:Firstly,a multi-view clustering method based on kernelization multi-view tensor analy-sis is proposed.In view of the poor performance of the previous algorithm in the nonlinear case,this thesis rewrites the optimization problem in multi-view subspace clustering,and maps the original data to the more linear high-dimensional subspace by kernel function,and then solves the self-representation coefficients in the high-dimensional subspace.For the problem after kernelization,this thesis proposes an effective optimization algorithm based on alternating direction iterative method.In order to further study the scalability of the multi-view tensor analysis clustering algorithm proposed in this thesis,this thesis extends the algorithm.Experiments show that the proposed algorithm improves the performance of multi-view clustering.The evaluation indicators on 8 cluster datasets including face,scene and general target exceed the current popular methods and even obtain a breakthrough im-provement.The average NMI and ACC reached 96.9%and 93.8%respectively.It is about 9.2%higher than the best methods in ACC.The convergence fastness and parameter robust-ness were also verified by experiments.Secondly,a multi-view index fusion method based on tensor analysis is proposed.Aim-ing at the problem that the previous index fusion method in image retrieval is difficult to explore the high-order information between the perspectives,this thesis applies the tensor k-ernel norm constraint to the index fusion optimization problem,and captures the high-order correlation of different perspectives in the unified tensor space without increasing the com-putational complexity.Experiments show that the proposed algorithm has comparable or even better performance than the popular methods.The evaluation indexes on the Holidays,Ukbench and Market-1501 datasets are 94.7%(mAp),3.49(NS-Score)and 50.92%(single query mAp)respectively.Thirdly,a deep clustering network based on deep learning is proposed.Aiming at the problem that the existing deep clustering algorithm is generally multi-stage and difficult to embed the supervised information.The deep clustering framework proposed in this thesis unifies the feature learning and clustering classification in image clustering to achieve end-to-end training,and joins supervision.The network branch enables a variety of different types of supervisory information to be embedded into the network through the branch to improve the performance of the cluster.Experiments show that the deep clustering algorithm with classification branches improves the performance of clustering.The clustering performance of dataset Sun-397 and self-built libraries ImageNet-100 and Caltech-256 increases 9.1%,10.0%and 5.8%respectively under the help of supervision information.