Unsupervised extraction of salient region-descriptors for content based image retrieval

Several content-based image retrieval systems use region-based descriptors of salient regions to represent the image content. In this application domain, no perfect pixel-wise segmentation is required, but instead, a stable and unsupervised extraction of salient region-descriptors is needed which might be generally applicable. Accordingly, in this paper an algorithm is proposed which combines local and area-based information of multidimensional features, such as luminance, color and texture to extract robustly region-descriptors of salient regions. Furthermore, the selection of the corresponding regions is parameter-free thus ensuring its general applicability. The algorithm is easy to extend to other feature types and has been used to extract salient regions from images of a large database consisting of outdoor scenes.