Research On Image Clustering Algorithm Based On Deep Unsupervised Learnin

Unsupervised learning is an important part of machine learning.The rapid development of deep learning has injected new vitality into unsupervised learning.Clustering is one of the basic problems in unsupervised learning,and the combination of deep unsupervised methods and clustering is the most promising development direction for clustering tasks in the future.This paper focuses on the exploration and extension of unsupervised learning in clustering tasks,proposes a mixture of adversarial autoencoders image clustering method,and carries out indepth research on this basis.The main research work is as follows:(1)We propose a method named a mixture of adversarial autoencoders image clustering network based on Generalized Gaussian distribution(MAAE).At present,a widely utilized assumption is that the data is a set of low-dimensional manifolds.So we represent each cluster of data by an adversarial autoencoder.The mixture network jointly learns the nonlinear data representation and the collection of adversarial autoencoders,The clustering results are obtained by minimizing the reconstruction loss of the mixture of adversarial autoencoders network.By introducing adversarial information,the hidden layer features of autoencoder can better match the prior target distribution of the Generalized Gaussian distribution.So the potential features of the data will be well separated.We have carried out experiments on standard datasets,and the results show that MAAE is effective in clustering and feature separation.(2)We propose a method named mixture of adversarial autoencoders image clustering network based on self-extraction of adversarial features(AMAAE).Considering that the artificially constructed target distribution has the limitation of low dimension,we add the adversarial features network to the network structure,and we introduce the mean center loss to train it.The input of the feature network comes from the sample selection of the mixture network,and the final clustering allocation is determined by both the minimum reconstruction loss and the mean center loss.Experiments show that AMAAE has been further improved in all aspects.(3)We propose a method named mixture of autoencoders image clustering network based on contrastive learning(CMAE).By introducing contrastive learning and the idea of “label is representation”,we add the contrastive learning network to the network structure.The whole network is divided into two parts,the mixture network only divides the positive and negative samples according to the reconstruction loss,it does not participate in the clustering allocation.The contrastive learning network carries on the contrastive learning on the soft-label vector of the positive and negative samples,and the final clustering allocation is directly completed by the soft-label vector.The experimental results show that the CMAE is effective in dividing positive and negative samples,feature learning and clustering.