A Statistical Comparative Study on Image Reconstruction and Clustering With Novel VAE Cost Function

Abstract

Deep clustering achieves unprecedented levels of accuracy with unsupervised feature extraction on rich datasets where the joint statistics of the latent space is learned via highly nonlinear compression. This paper has two separate contributions to this field. First, we conduct an extensive and first-of-its-kind empirical study on the statistical relationship between the clustering accuracy and image reconstruction quality of a state-of-the-art deep clustering topology in the form of a convolutional variational autoencoder (VAE) with a K-means back end. We change the latent variable z at the bottleneck of the network to create different latent dimensions and explore how clustering performance metrics and reconstruction metrics are statistically related. Secondly, based on our data-driven statistical findings, we also propose a novel cost function for the VAE which includes the structural similarity index measure to jointly optimize image quality and latent statistics for improved clustering. The preliminary results show significant increases in clustering accuracy of as much as 10.76% on two popular benchmark datasets. The TensorFlow implementation for the experimental framework can be found here: https://github.com/alla15747/IEEE-Comparitive-Study-VAE-Paper-(Python code will be available at the time of publication).

Keywords

• Auto-encoder (AE)
• computer vision
• convolutional neural network (CNN)
• deep learning (DL)
• image quality metrics
• variational auto-encoder (VAE)