Image classification using a stochastic model that reflects the internal structure of mixels

Abstract

This paper proposes new ideas for the classification of images with the presence of mixels, or mixed pixels. Based on the internal structure of mixels, we first propose a stochastic model called area proportion density, and we demonstrate that Beta distribution is an appropriate model for this density. Next, based on the linear model of a mixel, we derive another stochastic model called mixel density. This model is then incorporated into the mixture density model of the image histogram, and we show the peculiar flat shape of this model works particularly effective for image histograms with long tail. Finally we present experiments on satellite imagery, and the goodness-of-fit of the proposed model is evaluated from the viewpoint of information criterion.