Spatial Reasoning and Similarity Retrieval for Pictorial Databases Indexed by RCOS-strings with Shape Approximation

Abstract

Picture indexing and abstraction techniques are very important for pictorial database systems. Previous approaches, including 2D string, RCOS-string, ignore the shape information about objects and cause remarkable distortion in picture reconstruction, inaccuracy in spatial resoning, and ambiguity in similarity retrieval. In this paper, we present a spilt-and-merge method which will use a set of rectangles to approximate an object and adopt RCOS-string as the structure for iconic indexing. The spatial reasoning rules and a similarity retrieval algorithm based on this concept are also described. We built a prototype system and demonstrated that picture reconstruction, spatial reasoning and similarity retrieval become much more accurate than all previous approaches.