Relevance-based Content Modeling and Object Retrieval from Multi-source Image Data

Abstract

The problem of object retrieval for design automation based on semi-semantic representation of objects of interest in images is addressed in this article. The concept of an ordered set of salient feature vectors (SFVs) is introduced to concisely describe multi-source image data in different application areas. A system architecture is presented which combines statistical learning modules with multi-scale morphological modeling and analysis of image contents. In the presented approach, the object retrieval is based on establishing correspondence between two ordered sets of SFVs: a query reference image (or concise description of the object) and a database image. On a higher level, new rules of association are established between the design objects, based on the extracted SFVs and their spatial relations in images. Experiments with different types of images confirmed the utility of the proposed content modeling and proved the adequacy of the extraction accuracy of the SFVs.