Image sequence analysis using relational structures

摘要

Time varying images are usually analyzed to obtain a description of the observed objects and of their actions. An abstraction process is necessary which converts the input images into symbolic descriptions. The attributes of image symbols and the relationship between symbols are computed from the sampled intensity or colour measurements. Relational structures are a formal tool to describe not only the symbolic image representation but, furthermore, stored models as prototypes of objects and objects as instantiations of such prototypes. Following this paradigm, an essential task of image sequence analysis is the generation of mappings between images to establish the correspondence relationship and between images and prototypes to identify objects. Different types of morphisms between relational structures (RS-morphisms) are proposed, together with an approach for computing them by searching for cliques in a compatibility graph of hypothetical assignments of elements from both structures. The common formalism, presented here, aims at a unification of methods for those steps necessary to instantiate objects and follow them through the sequence. One practical reason for this attempt is to make (relational) database systems available to store and retrieve the large amount of data whidch might be generated during the exploration of a real-world image sequence.