A unified framework for retrieval in image databases
Venkat Naidu Gudivada · 1993
Image Retrieval (IR) problem is concerned with retrieving images that are relevant to users' requests from a large collection of images, referred to as the image database. Since the image database application areas are very diverse, the features of the existing image database systems have essentially evolved from domain specific considerations. In order to provide a generic approach to the IR problem, we have introduced a unified framework for retrieval in image databases. The framework supports four generic retrieval types: Retrieval by Browsing (RBR), Retrieval by Non-semantic Attributes (RNA), Retrieval by Spatial Constraints (RSC), and Retrieval by Semantic Attributes (RSA). To support these generic retrieval types, we have proposed an image data model referred to as AIR. AIR data model employs multiple logical representations. The logical representations can be viewed as abstractions of physical images at various levels and help toward achieving domain independence. We identify and provide solutions to the following three major research issues that arise in the context of AIR data model: algorithms for processing RSC queries, methods for the identification of semantic attributes, and algorithms for processing RSA queries. An algorithm, referred to as $SIM\sb{R}$, is developed for processing RSC queries. $SIM\sb{R}$ is robust in the sense that it can deal with translation, scale, and rotation variances in images. Personal Construct Theory (PCT) from clinical psychology is introduced as a knowledge elicitation tool for systematically deriving semantic attributes to support RSA. Also, an algorithm for RSA (based on the knowledge elicited from a domain expert using PCT) is developed. This algorithm incorporates user relevance judgments in the form of preference relations and incrementally improves retrieval quality.