Cognitive shape representation for the conceptual recognition and learning

Abstract

The cognitive shape representation that agrees with human intuition is proposed for the conceptual recognition and learning from pictogram or real shapes. Descriptions specifying the structure of shape in terms of meaningful parts and relations have cognitive power and anthropomorphism. A distinctive characteristic of this method is that, instead of using fixed computational representation, it describes the shapes by appropriate techniques depending on the shapes. Global shape analysis determines proper technique by interpreting various global features, such as collinearity, regularity, bending energy, compactness, and pattern ratio. The hybrid shape representation method that is organized hierarchically is presented for the conceptual recognition and fast matching with a large number of objects. Structural shapes are represented by cognitive shape decomposition and semantic network representation of parts and relations. For the cognitive shape decomposition that agrees with human intuition, many heuristic rules are applied on the context of collinearity, regularity, parallelism, and simplicity. Shape primitives are restricted to triangle, rectangle, and circle for reducing complexity. Each primitive is refined by similarity network. The structure of shape is described by the relationship among the primitives. From this cognitive description, classification of objects and concepts can be learned easily by generalization and specialization. A number of experiments conducted on the different types of shapes shows that the results correspond with human intuition.