A stochastic model of mixels and image classification

A pixel is not an “atom”. Most image processing algorithms basically assume that a pixel is not dividable; in other words, its inside is homogeneous consisting of only one classification class. This assumption is, however, not true for images, in particular remote sensing images, taken by a sensor with coarse resolution. In these images even a single image pixel may contain more data than two classification classes, which means that a pixel has an underlying heterogeneous structure internally. This paper discusses the statistical properties of these heterogeneous pixels, namely mixels. The difference between our work and previous work comes from the formulation of the problem-we focus on overall statistical properties that a set of pixels will show in the image intensity space. This formulation introduces our new stochastic model, mixel density, which also provides a proper interpretation for so-called long-tail density