A Direct Method for Semantic Partitioning of Low-level Image Data

Abstract

Abstract-It is a tough task to discover semantics implied by lowlevel image data. In view of this situation, single concept clustering (SCC), a new algorithm for semantic partitioning of data set according to a single concept is presented. First, data is preprocessed and an uniform interface obtained for the followup processing. Secondly, data set is described by one Gaussian， and all cases in which classes meet element gain or loss are classified into eight kinds. Three new theorems and two lemmas are established from the analyses of the eight cases. According to these theorems and lemmas, we combine the eight cases with the situations in which a class doesn't meet any element gain and loss, remove the relations between the previous class and the current class, form the relations between the current class and the succeeding class, and then draw four combinations. Finnally, data set is adaptively decomposed into semantic partitions with the four combinations. The experiments using the data of color images as the test data demonstrates that the SCC method can find sparse connected regions implying semantics, which lays a foundation for image label and analysis. Furthermore, the SCC method may also be used in other data processing tasks, for an example, determining equivalence classes of rough set.