2010 年 10 巻 1 号 p. 11-21
Families of t-norm based histogram intersections and a corresponding content-based image retrieval system are proposed. It is shown that the retrieval performance varies depending on the semantic category of the image and the usage of different t-norms results in a performance increase. Computational costs are determined as both the complexity and the average of 106 calculation times needed to calculate each similarity on a personal computer. To evaluate retrieval performance, precision and recall on the Corel image databases are determined, and Friedman test and Scheffe′s method are used to find difference between similarity measures. The results suggest that the proposals have the best average Friedman ranks in some categories (implying improved precision) having only 125 ms overhead when compared to the basic histogram intersection. Due to observed performance of families of t-norm based histogram intersection, system performing boosting of such norms could increase the overall accuracy in content-based image retrieval.