Image Clustering and Generation with HDGMVAE-I

Generating high-quality and realistic images has emerged as a popular research direction in the field of computer vision. We propose the Hierarchically Disentangled Gaussian Mixture Variational Autoencoder model with Importance sampling (HDGMVAE-I) for image generation and clustering. To enhance clustering performance, we introduce total correlation (TC) as a key to computing latent features between samples and cluster centroids and use slack variables to narrow down the feasible solution space. We also introduce Fisher discriminant as a regularization term to minimize within-class distance and maximize between-class distance, particularly effective for samples that are difficult to classify accurately. Moreover, we introduce importance sampling to reduce the information gap between the ELBO and the log-likelihood function, which leads to more realistic generated data. Experimental results demonstrate that our proposed method significantly outperforms the baseline in both clustering and generation tasks.