Convolutional Variational Autoencoders for Image Clustering

Abstract

The problem of data clustering is one of the most fundamental and well studied problems of unsupervised learning. Image clustering, refers to one of the most challenging specifications of clustering, concerning image data. Thankfully, the emerging Deep Neural Networks, and in particular Deep Autoencoders lead to the automation of image clustering, which until recently, was time consuming and labor intensive. However, the effect of the consideration of local structure during feature extraction from a Variational Autoencoder on clustering, is still an unstudied subject in the literature, while simultaneously constitute a baseline approach for supervised learning (Convolutional Neural Networks). For this reason, the methodology proposed in this paper, is composed from a Variational Autoencoder (VAE) surrounded by a convolutional network in a symmetric way. The resulting embedded image data are fed to various established clustering algorithms to examine clustering performance. In addition, we propose a modification of this approach, able to reduce complexity while achieving similar or even better clustering performance. Finally, we investigate the combination of VAE's produced embedding and manifold learning for image clustering. The extensive experimental analysis, verified the importance of the proposed methodology, exposing the potential for further developments. © 2021 IEEE.