Automatic extraction of image object definition from sample images

摘要

A method of automatic extraction of image object definition from sample images is reported. Object definition is a set of data including, the minimum bounding rectangle (MBR) of the object, a set of (attribute, MBR) pairs of components, and a set of neighbors of each component (called the neighbor set) representing the possible extension of the parent object seen from the component. Given the (attribute, MBR) pairs of segments, we first extract candidate object definitions out of grid windows placed on the image. Then we evaluate the "relevancy of object" of each candidate. The evaluation measure includes, number of instances of the object, the stability of component layout among instances, etc. Relevance of this approach are illustrated through experiments.