E. van den Broek
Digital imaging technology is more and more embedded in a broad domain. As a consequence digital image collections are emerging; this is no different for the cultural and historical domain. The need for efficient retrieval and data-mining in such collections eminent. For the latter purpose, content-based image retrieval (CBIR) techniques can be utilized. Such techniques exploit color, texture, shape, and spatial characteristics of the image content in their retrieval process. In the current research, we use CBIR techniques to label or retrieve the work of artists, without the need for complex annotations. Moreover, artists can be found of which the work resembles each other to a certain extend.
In general, for complex (multi-media) data-mining systems, user profiling is adopted as a paradigm. In contrast, we propose artist profiling instead. Subsequently, features as can be found in artists' work are identified (e.g., number of colors, contrasts, texture) and utilized in such a profile.
In this research, a range of topics are discussed and forthcoming problems are tackled; for example: How can human color and texture perception being modeled? Is it needed to incorporate relevance feedback algorithms and to what extend do people differ in their perception? What CBIR techniques are generally applicable and which compounds have to be adapted based on characteristics of a collection (e.g., paintings, photographs, statues).
This all will be illustrated in a demo version of CBIR system, based on artists' profiles.
|
