Research On Image Clustering Method Based On Variational Deep Embedding

Clustering can mine the underlying structure and patterns in a dataset and it is a widely used data statistical method nowadays.Traditional clustering algorithms have good performance when dealing with low-dimensional and small data sets.However,as the volume of data in a dataset becomes larger and the dimensionality of the data increases,it is difficult for traditional clustering algorithms to obtain better clustering results.In recent years,deep learning techniques have been rapidly developed and can be used to better learn the features of data for clustering.Researchers have proposed deep clustering algorithms that combine neural networks and clustering,by which use neural networks to extract the required feature dimension space from the original data samples and transform high-dimensional data into low-dimensional features,which helps to obtain better clustering results.Clustering based on deep learning is currently the mainstream method,but there is still room for improvement in feature extraction.Designing more effective deep clustering neural network structures and conducting deep clustering research that can extract higher quality features have important theoretical and practical significance.The research on the deep clustering algorithm includes the following two aspects:(1)A convolutional variational self-encoder-based deep clustering algorithm is constructed.Based on Variational Deep Embedding(Va DE),the model firstly processes the original image data by data augmentation to increase the diversity of data samples and improve the generalization ability of the model,so that the original data can be learned better and the final clustering results can be improved.Furthermore,a convolutional layer is added to the variational self-encoder to build a convolutional variational self-encoder.The encoder of the final convolutional variational self-encoder consists of two convolutional layers and one fully-connected layer,and the decoder consists of one fully-connected layer and three deconvolutional layers.The features are extracted from the original data samples using the mirror structure of the convolutional variational self-encoder,and then the network is jointly trained and fine-tuned with feature learning and clustering.The clustering accuracy in the experiments conducted on the publicly available image datasets FASHION-MNIST and STL-10 is 62.01% and86.47%,respectively,which is 3.88% and 1.97% better than the clustering accuracy of the original variational deep embedding algorithm,respectively.(2)A deep clustering algorithm based on the CBAM attention mechanism is constructed.The convolutional block attention module(CBAM)can distinguish the relevant and irrelevant information features,by which the model can determine the dynamic weight parameters to reinforce the key information and weaken the useless information.Thus,the model is able to capture the key features of the input image and improve the clustering results.To verify the effectiveness of the proposed deep clustering algorithm,experiments were designed on two publicly available image datasets,FASHION-MNIST and STL-10.The clustering accuracy was further improved to 64.67% and 88.18%,which increased by 6.54% and 3.68% compared to Va DE,respectively.