Statistical modeling of multi-camera images

Abstract

This paper aims at developing mutual statistical descriptions of images in a visual sensor network. A two-dimensional Gaussian scale mixture (GSM) model is proposed to describe the joint distribution of wavelet coefficients in the same subband of two different images. This model is tested on two image databases, and it is shown that the proposed model always achieves a smaller divergence from the empirical distribution than a (jointly) Gaussian model. The GSM model is then extended to blocks of wavelet coefficients in the same subband and scale of non-overlapping images. A new denoising algorithm is devised that uses this model and achieves better performance than denoising based on marginal models.