Exploring Region Correlation and Context Expansion for Image Retrieval

Xinjing Wang,Wei‐Ying Ma,Xing Li

2005-01-01

Abstract:ABSTRACT We explore the statistical correlations between homogeneous regions after image segmentation for improving image retrieval performance. Our proposed approach, called Context Expansion , is analogous to query expansion in text retrieval where the statistical correlations between words are used to expand query for improving search precision and recall. In our approach, regions in images are classified into two types: key region representing the main semantic content and environmental region representing the context. We perform context expansion for the key region in the query example using highly correlated environmental regions. In our experiments, we found that this method is able to filter irrelevant images (i.e. improving precision) as well as retrieve relevant images whose context may be varied (i.e. improving recall). We also study several major factors which have impact on the performance of context expansion. Categories and Subject Descriptors H.3.3 [ Information Search and Retrieval