The purpose of this paper is to realize a fast image classification method with considering the effect of mixels by extending maximum likelihood thresholding methods. A ``mixel'' is generated from the process of several classes being mixed within a single pixel. Here this paper hypothesizes that a population of mixels generates a hidden distribution called a ``mixel distribution'' on the gray-level histogram. Moreover, the discovery of the fact that the moments of this mixel distribution can be derived analytically leads to the proposition of a new thresholding method based on the definition of a likelihood criterion that takes the effect of mixels into consideration and the formulation of the algorithm for maximizing the proposed likelihood criterion. The experiments on scanned imagery and satellite imagery demonstrates that the proposed method is successful in selecting thresholds on gray-level histogram for better discriminating between peak and long-tail regions which are presumably produced by mixels. In terms of speed, the proposed method achieves considerably faster speed than conventional methods.

      @InProceedings{ k:prmu99,
      author = {Asanobu KITAMOTO},
      title = {A Maximum Likelihood Thresholding Method Considering the Effect of Mixels},
      booktitle = {Technical Report of Institute of Electronics, Information, and Communication Engineers (IEICE)},
      volume = {PRMU99-166},
      pages = {7-14},
      year = 1999,
      month = 12,
      doi = {},
      note = { (in Japanese)},
      }