Image segmentation via minimal-length encoding
Y.G. Leclerc · 2003
Summary form only given. Vision is treated as an inference process. The basic motivation behind this inference process is the hypothesis that if a language that provides an efficient description of a large number of observations (images) can be found, then the simplest descriptions in that language tell something about the causes of the observations. It is proposed to achieve this simplest possible description in a hierarchical fashion, where each level is itself posed as a minimal-length encoding problem. The key idea of the hierarchy is that incrementally more efficient descriptions can be obtained by an incremental decomposition of the image into ever smaller groups of causal processes. At the lowest level (the image), all of the causal processes are grouped into a single description as an array of intensities.>