QUALITATIVE SPATIAL RELATION REPRESENTATION AND RETRIEVAL

Abstract

Pictorial databases become more and more popular recently. However, in pictorial databases, images of pictures always require large storage and are not easy to be retrieved. Therefore, "content-based" databases retrieval attracts more and more attentions in these years. Rather than proceeding in the original full-sized images, content-based retrieval works by matching the query against a representation of the content of the image. This representation of the image is a symbolic version of the image which will be used to represent the objects in the picture and features extracted from the original picture. Normally, the features of the corresponding pictures include two forms: visual features (such as color, texture, and shape) and relation features (spatial relations among the objects in a picture). In this thesis, we analyze two dimension (2D) space spatial relation representation and similarity retrieval. We introduce some typical qualitative spatial relation representations which include using two one dimension (ID) relations to represent 2D relations, using one integrated 2D relation to represent and index 2D pictures, and using systematic topological and orientation relations to represent 2D space relations. A variety of retrieval mechanisms for picture similarity retrieval is also presented. In particular, we propose a qualitative spatial relation representation which is called Intrinsic Orientation and Topological relation representation (IO&T). This representation combines topological relation and orientation relation, and applies fuzzy concept to categorize intrinsic orientation which achieves rotation invariant characteristic and overcomes other anomalies caused by human's fuzzy perceptions in spatial relation representation. A new object based retrieval Qualitative spatial relation representation and retrieval approach is also proposed based on this representation. By building a topological and orientation relation similarity measure system, this retrieval approach makes use of some objects which are dominant objects in the query for retrievals. These dominant objects are actually those objects which are most concerned in queries. They should be defined by users, and hence make the similarity retrieval user oriented. Experiments and prototyping model are used to demonstrate the feasibility of our proposed representation and retrieval approaches. The results show our representation can capture more accurate features of pictures, such as rotation and real physical relations, and is also more intuitive than other representations. By trimming the answer set of similar pictures effectively using our defined IO&T similarity measure and user defined dominant objects, the proposed object based retrieval is also proven to be more practical because its complexity is not too high due to the use of only few dominant objects, and more complete because it can always cover the real answer set.