We offer here a non-conventional approach to muhivariate ima- structured
data for which the basis is well tested but the analytical ramifi-
cations are still unfolding. Although we do not formally pursue them,
there are several parallels with the nature of neural networks. We
employ a systematic set of statistical heuristics for modeling
multivariate image data in a quasi-perceptual manner. When the human eye
perceives a scene, the elements of the scene are segregated
heuristically into compo- nents according to similarity and
dissimilarity, and then the relationships among the components are
interpreted. Similarly, we segregate or seg- ment the scene into
hierarchically organized components that are subject to subsequent
statistical analysis in many modes for interpretive purposes. We refer
to the segregated scene segments as patterns, since they provide a basis
for perception of pattern. Since they are also hierarchically organ-
ized, we refer to them further as polypatterns. This leads us to our
acro- nym of Progressively Segmented Image Modeling As Poly-Patterns
(PSIMAPP). Likewise, we formalize our approach in terms of pattern
processes and segmentation sequences. In alignment with the terminology
of image analysis, we refer to our multivariate measures as being signal
bands.