COALESCE: A Probabilistic Ontology-Based Scene Understanding Approach1

Abstract – An important component of higher level fusion and decision making is knowledge discovery. One form of knowledge representation is a set of probabilistic relationships between entities. Here we present biologically-inspired algorithmic support for automatic scene understanding and complex object recognition. Our algorithm learns the association between scene and complex objects and their primitive components with and/or without a priori knowledge. In addition, the spatial relationships between the simple constituents and their probabilities are learned incrementally. Complex Object Associative Learning Enables SCene Exploitation neural network (COALESCE) is a hybrid neural network based on probabilistic associative learning and hyper-elliptical learning algorithms. The Object Probabilistic Associative Learning (OPAL) algorithm automatically discovers the conditional probabilities and hierarchical structure comprising a scene. Hyper-Elliptical Learning and Matching (HELM) learns spatial relationships between objects in an object-centric reference frame