Overlapped latent Dirichlet allocation for efficient image segmentation

Abstract

Unsupervised methods for image segmentation have recently drawn significant attention because most images do not have labels or tags. A topic model is an unsupervised probabilistic method that captures the latent aspects of data, where each latent aspect or topic is associated with one homogeneous region. In this paper, we propose a new topic model for image segmentation task that incorporates spatial information into its structure based on the hypothesis that overlapped topic proportions convey spatial information. The model is efficient in time and memory, and we demonstrate this through comparison with other models using the MSRC image dataset.