基于变分自编码器与流形特征的聚类算法OACSTPCD

Deep neural network has become a promising clustering method due to its excellent nonlinear mapping ability and flexibil-ity in different scenarios.In order to map the original high-dimensional data to a feature space in where the clustering is easy to be done,feature extraction or feature transformationare are done by many deep clustering methods,and then the extracted features are grouped into different clusters in the lower-dimensional space,which still are assumed in Euclidean space.In order to explore the im-pact of feature extraction and manifold space on clustering performance,in this paper,we propose a clustering algorithm based on variational autoencoder and manifold learning—MFVC(Clustering Algorithm Based on Variational Autoencoder and Manifold Fea-tures).In this method,the β-VAE(Learning Basic Visual Concepts with a Constrained Variational Framework)with residual connec-tion layer is used as a feature extractor to extract image features,and the non-parameter attention mechanism SimAM(A Simple,Pa-rameter-Free Attention Module for Convolutional Neural Networks)is added to improve the expressive ability of the convolutional network.For more favorable features,the Manifold UMAP(Uniform Manifold Approximation and Projection for Dimension Reduc-tion)method is used to improve the separability of the features,and then the K-Means method is used for clustering learning.Experi-mental results on six benchmark datasets show that this method can provide better performance.MFVC achieves with accuracy of 0.981 on the MNIST(Mixed NationalInstitute of Standards and Technology database)dataset,and 0.681 on the Fashion-MNIST da-taset.