<title>Evaluation of contents-based image retrieval methods for a database of logos on drug tablets</title>

In this research an evaluation has been made of the different ways of contents based image retrieval of logos of drug tablets. On a database of 432 illicitly produced tablets (mostly containing MDMA), we have compared different retrieval methods. Two of these methods were available from commercial packages, QBIC and Imatch, where the implementation of the contents based image retrieval methods are not exactly known. We compared the results for this database with the MPEG-7 shape comparison methods, which are the contour-shape, bounding box and region-based shape methods. In addition, we have tested the log polar method that is available from our own research. For the evaluation of the rotation invariance and general effectiveness of the image retrieval method we used three different tablets, which were digitized in 25 different positions. From our experiments it appeared that log polar correlation worked well (relevant matches in top positions) for the shape of the logo. However, the problem with this approach is that it takes a huge computing power, and that it is not capable of making fast search indexes based on descriptors of features. The MPEG-7 methods for this database did not deliver all relevant matches in top positions. The QBIC and Imatch software gave better results due to the influence of their effective color descriptors. With one stamp more colors of pills can be manufactured, so comparing on color is not relevant for the shape comparison of the logos. With the MPEG-7 the results improved after preprocessing and segmenting the logo from the background of the pill. After this processing, the contour-shape based method gave similar results as the log polar correlation.