| Clustering analysis is a kind of vital methods in data mining,which can find of potential knowledge and rules from large-scale unlabelled data.It has been widely used in pattern recognition,image processing,machine vision,and so on.There are required to analyze and handle image data in many high-technology applications such as driverless,tracking and localization,as well face recognition.Some conventional clustering algorithms cannot effectively deal with high-dimensional complex image data.Thus,how do efficiently and effectively group high-dimensional complex image data into clusters remains a challenging problem,and these further researching results are of great theoretic significance and practical values.This research focused on unsupervised clustering analysis for high-dimensional complex image data.Motivated by the advantages of deep features representation techniques,two deep clustering algorithms were proposed to improve clustering performance.First of all,an AEMDPC algorithm based on fully connected AutoEncoder network combined with density peaks clustering method was proposed.Secondly,an alternative deep embedding clustering DDC algorithm was proposed jointly learning features representation and partitioning clusters.Main ideas of the two methods are described as follows:1.To improve the worse clustering performance of DPC algorithm when processing complex distributed data and high-dimensional image data,an AE-MDPC algorithm was proposed using the deep network combined with manifold distance.The Gaussian function value of manifold distance was used as local density to describe the global consistence,which leads to the local density of potential centers growing local peaks and then was selected as the cluster centers.The fully connnected AutoEncoder learned the optimal features representation and simultaneously removed feature noises.The AE-MDPC algorithm improved the separation of the density values,reduced the "domino effect" of the DPC algorithm as well.It offers better clustering performance on seven UCI and image datasets than the DPC algorithm.2.AE-MDPC algorithm separately performed features extraction and clustering analysis,which significantly influences the clustering performance.To this end,another deep clustering method was proposed(referring to DDC),which employed a deep denoising convolutional AutoEncoder to learn robust features representation from noisy image data,and combined with self-attention mechanism to obtain the critically local features.End-to-end jointly training was employed to fine-tune the network parameters that bring more suitable features representation to clustering tasks and then further refine clustering results.The proposed DDC clustering method was evaluated on several popular image datasets.The experimental results demonstrated that DDC provide better clustering performances and outperformes the state-ofthe-art deep clustering algorithms... |
