Similarity Evaluation based on Image Primitives

S. Scholz (Germany)


shape, content based image retrieval, trademark images


A new framework for the perceptually relevant compari son of figurative images, especially trademark logos is pre sented in this paper. Images are divided into salient ge ometric figures such as rectangles, ellipses, and triangles. Parts not fitting into any of those simple classes are rep resented by their boundaries. The figures are classified, re lated, and weighted according to their perceptual relevance. For the comparison of two images the figures and the re lations are compared independently from each other. For the comparison of single figures a simple measure of sim ilarity based on registration techniques is applied, which is noise tolerant and shows good results for figurative im ages that have no spatially independent parts. The simi larity of the images is then determined by the similarities of the figures and the relations for the best match. The al gorithms were tested with a collection of 10 745 trademark images from the UK PTO, with the same set of 24 reference queries that were used to test the ARTISAN System. Each query consists of a query image and a list of relevant im ages, compiled by experienced trademark examiners. The experiments show that the presented approach allows for a considerable improvement of content based image retrieval in trademark images.

Important Links:

Go Back