Abstract
A comprehensive comparison of the three variants of region-based image similarity models employing the Earth Mover's Distance (EMD) is performed. An important characteristic of the EMD - that makes it particularly suitable for the region-based image similarity modeling- is that it allows for comparison of images with different number of regions. Despite this fact, very few of the existing region-based models employ EMD. Common for the proposed three variants are
the image features and the employment of fuzzy aggregation operators, while the difference is in the way elements of the EMD - flow matrix and feature distributions - are modeled. Empirical evaluation of the retrieval performance of the three variants on four test databases, containing 4444 general-purpose images in 150 semantic categories is done. Comparison of the proposed variants to the representative of the existing region based models is done, too. Results suggest, that proposed models result in different retrieval performance and suggest possible improvements that can be achieved along the same line of research directions.
