As we mentioned before, the different types of queries and their relative frequency will greatly influence the way images (or their compressed versions) should be organized in storage devices to minimize the browsing and retrieval time in future Digital Library Systems. For example, consider the following scenario. A user retrieves an image at a low resolution and then requests a sequence of one or more higher resolution subimages of the original image. Each time the subimages become smaller as the user narrows down the search. Perhaps these images (or subimages) would be simultaneously viewed in one or more windows of several screens located at different sites. In order to serve a reasonably large community of users efficiently, we not only need a parallel computer system with access to multiple storage devices, we also need efficient algorithms to process the data, and an efficient way to store and retrieve subimages. This type of query could be followed by one in which the same subregions in other images are searched for the presence of the same type of feature encountered in the original search, and the images containing such features would then be displayed at several different sites. This additional search could be automatic, semi-automatic, or manual depending on the specifics of the search, and in some applications, the search could be aided with searches carried out in additional databases and catalogues.
Images need to be stored in such a way that multiresolution images can be efficiently retrieved, and so that one can quickly retrieve only portions of these images without having to decompress the whole image. This extraction of subimages needs to be performed in parallel in order for the system to have a fast response time. One can develop several storage and compression strategies that could be efficiently used in a parallel computer system with multiple storage devices. For example, to allow multiple simultaneous accesses to different portions of an image, one can hierarchically partition the image. A quadtree representation [44] or a more general partition [6] can be constructed where each terminal node represents an image block of relatively homogeneous composition. Furthermore, a decomposition can be performed on individual image blocks instead of on the original image to construct multiple, multi-resolution representations of the image. This scheme allows concurrent decomposition and access to different portions of an image and therefore is suitable for implementation on a parallel computer system. The selection of an appropriate access and compression strategy would depend on the type of queries and their relative frequency.