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 probabilistic 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 probabilistic model called mixel density. This model is then incorporated into the mixture density model that describes the image histogram and we suggest that the peculiar flat shape of this density effectively models long-tail image histograms. Finally we present experiments on satellite imagery, and the goodness-of-fit of the proposed models is evaluated from the viewpoint of information criterion.

      @InProceedings{ kt:miru98,
      author = {Asanobu KITAMOTO and Mikio TAKAGI},
      title = {Image Classification Method Using Probabilistic Models that Reflect the Internal Structure of Mixels},
      booktitle = {Meeting on Image Recognition and Understanding (MIRU'98)},
      volume = {I},
      pages = {87-92},
      year = 1998,
      month = 7,
      doi = {},
      note = { (in Japanese)},
      }