Dynamic Distance Metric for Image Retrieval Systems

Query-by-one-example (QBE) has been a popular query system for content-based image retrieval (CBIR) for more than a decade. However, recent research has shown that a single image is not sufficient to form its semantics or concept of the intended query. Searching concept “car,” for instance, one might need many examples of car images in various colors. The color feature is then understood as a non-factor in the distance metric. This paper proposes a novel approach, which users can query using groups of query images. There are three possible groups: relevant (positive), irrelevant (negative) or neutral groups. The range for each feature within these groups of query images is defined. These ranges are used to adjust the weights of the features. As a result, some features may be cancelled out from the similarity computation. The measure then becomes a dynamic metric for image retrieval. This novel approach achieves a higher degree of precision and recall and, at the same time, significantly reduces the time complexity of matching. The proposed approach is tested against the ImageGrouper method. The results show that this approach is an effective and efficient technique for image retrieval systems.

Keywords: dynamic distance metric, range distance, content-based image retrieval, query-by-example

Corresponding author: E-mail: khnualsa@kmitl.ac.th , soonthar@cs.ucf.edu