Context-sensitive processing of semantic queries in an image database system

摘要

In an image database environment, an image can be retrieved using common names (labels) of entities that appear in it (such as door, book, car, …, etc.). In many cases, we specify further details of these entities and some relations between them. A semantic query is the formal expression method of label-based image retrieval requests. This paper shows how an image is abstracted into a hierarchy of entity names and features (such as brightness, length, …, etc.), and how relations are established between entities visible in the image. Semantic queries are also hierarchical. However, they are short and often overlook certain levels of the abstraction of the requested image. The core of this paper is a fuzzy matching technique that compares semantic queries to image abstractions by assessing the similarity of contexts between the query and the candidate image. An important objective of this matching technique is to distinguish between abstractions of different images that have the same labels but are different in context from each other. Each image is tagged with a matching degree (against the query) even when it does not provide an exact match of the query. Several experiments have been conducted to evaluate the strategy presented in this paper.