Appearance-Based Global Similarity Retrieval of Images
Sai Ravela, Can Luo · Kluwer Academic Publishers eBooks · 2005
Visual appearance is an important part of judging image similarity. We readily classify objects that share a visual appearance as similar, and reject those that do not. Our hypothesis is that image intensity surface features can be used to compute appearance similarity. In the first part of this paper, a technique to compute global appearance similarity is described. Images are filtered with Gaussian derivatives to compute two features, namely, local curvatures and orientation. Global image similarity is deduced by comparing distributions of these features. This technique is evaluated on a heterogeneous collection of 1600 images. The results support the hypothesis in that images similar in appearance are ranked close together. In the second part of this paper, appearance-based retrieval is applied to trademarks. Trademarks are generally binary images containing a single mark against a texture-less background. While moments have been proposed as a representation, we find that appearance-based retrieval yields better results. Two small databases containing 2,345 parametrically generated shapes, and 10,745 trademarks are used for evaluation. A retrieval system that combines a trademark database containing 68,000 binary images with textual information is discussed. Text and appearance features are jointly (or independently) queried to retrieve images. These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.